Method for self-adapting accurate measurement of dry bulk cargo of multiple types based on three-dimensional radar point cloud

Abstract

The application discloses a dry bulk cargo multi-species adaptive precision metering method based on three-dimensional radar point cloud, and relates to the technical field of intelligent metering of port dry bulk cargo. The method comprises the following steps: acquiring a multi-source collection data adaptive metering parameter set of a dry bulk cargo metering area; determining a ship body swing level according to ship body posture data and controlling a servo stability augmentation holder to execute reverse posture compensation; determining a dust interference level according to a point cloud quality index, and adaptively adjusting point cloud collection parameters and processing parameters in combination with the ship body swing level; pre-calculating a combined coordinate transformation matrix according to a time slice, and unifying point clouds of different visual angles to a same reference coordinate system; then determining a fine processing area and a simplified processing area through spatial grid quality grading; and finally performing adaptive curved surface reconstruction, volume calculation, weight conversion and outputting metering confidence. The application can improve the accuracy, stability and processing efficiency of dry bulk cargo metering under the scenes of dust interference, ship body swing and multi-species switching.

Description

Technical Field

[0001] This invention relates to the field of intelligent metering technology for dry bulk cargo in ports, specifically to an adaptive and accurate metering method for multiple types of dry bulk cargo based on three-dimensional radar point clouds. Background Technology

[0002] Port dry bulk cargo typically includes bulk materials such as coal, ore, mineral powder, coal powder, sand and gravel, and grain. Existing methods for measuring dry bulk cargo at ports mainly include manual measurement, weighbridge weighing, fixed single-point radar ranging, and volume calculation based on 3D point clouds. Manual measurement suffers from high operational risks, high labor intensity, and significant subjective errors; weighbridge weighing relies on vehicles or loading / unloading processes, making it difficult to achieve real-time dynamic inventory checks in ship holds and storage yards; fixed single-point radar can only obtain local height information, making it difficult to reconstruct the complete surface of the stockpile. 3D radar point cloud measurement can obtain a 3D point cloud of the dry bulk cargo stockpile surface through non-contact scanning, and further perform surface reconstruction and volume calculation, offering advantages such as high automation, low operational risks, and wide applicability. However, in the actual port operating environment, existing 3D radar point cloud measurement solutions still have the following shortcomings.

[0003] Different dry bulk cargoes exhibit significant differences in particle size, bulk density, angle of repose, surface roughness, and radar reflection characteristics. Powdered materials such as coal powder and mineral powder are prone to dust generation during loading and unloading, leading to numerous outliers, reflection gaps, and voids in the point cloud. The surface of massive ores is highly uneven, resulting in significant local curvature in the actual stockpile profile. Using fixed filtering parameters can easily lead to residual dust noise or over-smoothing of the actual profile. Port open-air storage yards and ship loading / unloading areas often experience complex environments such as high dust levels, fog, rain, snow, strong light, and alternating night and day conditions. Dust particles can cause anomalies in 3D radar echoes, resulting in outliers, local gaps, and edge breaks in the point cloud. Traditional pass-through filtering, statistical filtering, or fixed threshold filtering typically only handles small-scale noise. When dust causes large-area point cloud gaps, simple linear interpolation is insufficient to recover a dry bulk cargo stockpile surface that conforms to the actual packing patterns. In shipboard metering scenarios, the ship's roll, pitch, bow roll, and heave cause the coordinate system of the shipborne 3D radar to change with the ship's movement, resulting in single-frame point cloud scanning distortion, multi-frame point cloud registration offset, material pile edge drift, and point cloud ghosting. If only backend algorithms are relied upon for global point cloud registration and attitude compensation, the computational load is large, the real-time performance is poor, and cumulative errors are easily generated when strong shaking and high dust levels occur simultaneously.

[0004] Furthermore, existing methods typically perform filtering, registration, completion, and reconstruction processes of uniform complexity on the entire point cloud. For large, flat, and stable point cloud regions, performing high-complexity reliability assessments and high-density mesh reconstruction point by point results in computational redundancy. For localized dusty areas, point cloud voids, material pile edges, and highly undulating areas, even finer processing is required to ensure measurement accuracy. Existing solutions lack a data processing mechanism that prioritizes quality grading followed by local fine processing, leading to insufficient real-time measurement efficiency at the port. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive and accurate measurement method for multiple types of dry bulk cargo based on three-dimensional radar point clouds, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds, comprising the following steps:

[0007] S1. Acquire multi-source data from the dry bulk cargo metering area and determine the adaptive metering parameters corresponding to the current cargo type based on the cargo type information.

[0008] S2. Determine the ship's sway level based on multi-source acquired data, and control the servo stabilization gimbal to perform reverse attitude compensation for the shipborne three-dimensional radar.

[0009] S3. Determine the dust interference level based on the point cloud quality index of multi-source data, and adaptively adjust the point cloud acquisition parameters and processing parameters in conjunction with the ship sway level and cargo type adaptive metering parameters.

[0010] S4. Divide the dry bulk cargo point cloud into multiple time slices according to the multi-source acquisition data, and pre-calculate the combined coordinate transformation matrix for residual attitude correction for each time slice.

[0011] S5. Based on the combined coordinate transformation matrix, the dry bulk cargo point clouds from different acquisition perspectives are uniformly transformed to the same reference coordinate system to obtain the correction point cloud;

[0012] S6. Divide the corrected point cloud into multiple spatial grid cells and determine the grid quality level based on the point cloud quality characteristics of each spatial grid cell.

[0013] S7. Determine the fine processing region and the simplified processing region based on the raster quality level. Perform point-level confidence assessment on the fine processing region and perform adaptive voxelization downsampling and lightweight filtering on the simplified processing region.

[0014] S8. Based on the point-level credibility assessment results and the adaptive metering parameters of the cargo type, suppress dust outliers, retain real pile surface points and suspected boundary points, and only perform physical constraint completion on the missing area of ​​the point cloud and its adjacent area to obtain the completed pile surface point cloud.

[0015] S9. Based on the completed point cloud of the stack surface, the adaptive metering parameters of the cargo type, and the degree of local stack surface undulation, adaptive surface reconstruction and volume calculation are performed to obtain the volume measurement results of dry bulk cargo.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] (1) This invention generates corresponding adaptive measurement parameters for cargo by introducing cargo type information, so that dry bulk cargoes with different particle sizes, surface roughness, stacking morphology and radar reflection characteristics can adopt point cloud acquisition, filtering, dust suppression, completion and surface reconstruction parameters that match their cargo type characteristics, thereby reducing the probability of dust noise residue and excessive smoothing of the real stack surface profile under fixed parameter conditions; by controlling the servo stabilization gimbal to perform reverse attitude compensation for the shipborne three-dimensional radar according to the ship sway level, and further pre-calculating the combined coordinate transformation matrix according to the time slice, the residual attitude error after mechanical stabilization is corrected by coordinate correction, which can reduce the impact of ship roll, pitch and yaw on point cloud scanning attitude, reduce misalignment, ghosting and material pile edge drift between multiple frames of point cloud, and reduce the computational burden caused by relying entirely on back-end global registration.

[0018] (2) The present invention determines the dust interference level based on the point cloud quality index, and dynamically adjusts the point cloud acquisition parameters and processing parameters in combination with the ship sway level and cargo type adaptive measurement parameters, so that the point cloud processing process can adjust the parameters according to different dust interference states and ship motion states.

[0019] (3) This invention divides the corrected point cloud into spatial grid units and determines the grid quality level. It performs point-level reliability assessment on low-quality areas and performs adaptive voxelization downsampling and lightweight filtering on high-quality areas. This allows the computational resources to be concentrated on dust interference, missing point cloud, material pile edges and local undulation areas, reducing computational redundancy caused by repeated high-complexity processing on stable quality areas. Furthermore, it combines point-level reliability and cargo type adaptive metering parameters to remove dust outliers, retain real pile surface points and suspected boundary points, and performs local completion only on the missing point cloud areas and their adjacent areas according to physical constraints. This avoids large-scale interpolation of the complete measured pile surface, which changes the real geometric shape. Finally, it adjusts the surface reconstruction scale and completes the volume calculation according to the cargo type adaptive metering parameters and the degree of local pile surface undulation. This takes into account the dust suppression requirements of powdery materials, the local contour preservation requirements of blocky materials and the attitude correction requirements under shipborne dynamic metering conditions, thereby improving the adaptability, real-time processing capability and volume reconstruction accuracy of dry bulk cargo three-dimensional point cloud volume measurement. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the present invention.

[0022] Figure 2 This is a timing diagram showing the interaction between the modules of this invention.

[0023] Figure 3 This is a timing diagram showing the interaction between hull sway compensation and point cloud acquisition in this invention.

[0024] Figure 4 This is a flowchart of the time slice combination coordinate transformation of the present invention.

[0025] Figure 5 The diagram shows a comparison between the point cloud results of the processing method of the present invention and the prior art. (a) is a schematic diagram of the original point cloud result of the prior art or without sufficient optimization processing, and (b) is a schematic diagram of the completed stacked surface point cloud result after processing by the method of the present invention.

[0026] Figure 6 This is a bar chart comparing the volume measurement errors of the three methods of this invention.

[0027] Figure 7 This is a dumbbell diagram comparing the actual pile surface point retention rates under dust interference conditions according to the present invention.

[0028] Figure 8 This is a bar chart showing the comparison of measurement errors for multiple product types in this invention.

[0029] Figure 9 This is a line graph showing the processing time of this invention as a function of point cloud size. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-9 As shown, the present invention provides a technical solution: an adaptive and accurate measurement method for multiple types of dry bulk cargo based on three-dimensional radar point clouds, comprising the following steps:

[0032] S1. Acquire multi-source data from the dry bulk cargo metering area and determine the adaptive metering parameters corresponding to the current cargo type based on the cargo type information.

[0033] S2. Determine the ship's sway level based on multi-source acquired data, and control the servo stabilization gimbal to perform reverse attitude compensation for the shipborne three-dimensional radar.

[0034] S3. Determine the dust interference level based on the point cloud quality index of multi-source data, and adaptively adjust the point cloud acquisition parameters and processing parameters in conjunction with the ship sway level and cargo type adaptive metering parameters.

[0035] S4. Divide the dry bulk cargo point cloud into multiple time slices according to the multi-source acquisition data, and pre-calculate the combined coordinate transformation matrix for residual attitude correction for each time slice.

[0036] S5. Based on the combined coordinate transformation matrix, the dry bulk cargo point clouds from different acquisition perspectives are uniformly transformed to the same reference coordinate system to obtain the correction point cloud;

[0037] S6. Divide the corrected point cloud into multiple spatial grid cells and determine the grid quality level based on the point cloud quality characteristics of each spatial grid cell.

[0038] S7. Determine the fine processing region and the simplified processing region based on the raster quality level. Perform point-level confidence assessment on the fine processing region and perform adaptive voxelization downsampling and lightweight filtering on the simplified processing region.

[0039] S8. Based on the point-level credibility assessment results and the adaptive metering parameters of the cargo type, suppress dust outliers, retain real pile surface points and suspected boundary points, and only perform physical constraint completion on the missing area of ​​the point cloud and its adjacent area to obtain the completed pile surface point cloud.

[0040] S9. Based on the completed point cloud of the stack surface, the adaptive metering parameters of the cargo type, and the degree of local stack surface undulation, adaptive surface reconstruction and volume calculation are performed to obtain the volume measurement results of dry bulk cargo.

[0041] In the above implementation, step one is first performed to collect multi-source data on the dry bulk cargo metering area. This multi-source data includes dry bulk cargo point cloud data, cargo type information, and ship attitude data. Specifically, the dry bulk cargo point cloud data is collected by shipborne 3D radar and is used to characterize the 3D spatial position of the dry bulk cargo stack. The cargo type information is used to determine the cargo category of the current metering object. The ship attitude data is used to characterize the ship's roll, pitch, and bow states during the point cloud acquisition process. All types of data are recorded using a unified time reference, and a correspondence is established based on the acquisition time, so that the point cloud data, cargo type information, and ship attitude data collected within the same time period form a set of associated data. After obtaining the cargo type information, the adaptive metering parameters corresponding to the current cargo type are read according to the pre-established correspondence between cargo type and metering parameters. Each cargo type is configured with an independent parameter group, which specifies the processing conditions for that cargo type during point cloud acquisition, point cloud quality determination, dust outlier identification, voxel downsampling, point cloud completion, and surface reconstruction. When the cargo information changes, the current parameter group is switched to the parameter group corresponding to the changed cargo, so that subsequent processing uses the same parameter configuration as the current cargo.

[0042] In the above implementation, step two involves determining the hull sway level based on the hull attitude data. Within a pre-set time window, the roll, pitch, and bow angles are continuously read, and the values ​​of each attitude angle at the start and end of the time window are recorded. For each attitude angle, the attitude angle value at the end of the time window is subtracted from the attitude angle value at the start of the time window, and the absolute value of the result is taken to obtain the corresponding attitude angle change. Simultaneously, the attitude change rate is calculated based on the attitude angle change between two consecutive sampling times and the sampling time interval. The roll, pitch, and bow angle changes, along with the corresponding attitude change rate, are compared with pre-set grading thresholds, and the hull sway level is output according to a pre-established combination judgment rule. For example, if all attitude parameters are within the first threshold range, it is determined to be the first sway level; if any attitude parameter enters the second threshold range, it is determined to be the second sway level; and if any attitude parameter reaches the third threshold range, it is determined to be the third sway level. Each hull sway level corresponds to a pre-defined set of servo compensation parameters, time slice length parameters, and residual attitude correction parameters. Therefore, once the hull sway level is determined, the corresponding parameters can be directly called. After determining the hull sway level, the reverse attitude compensation amount of the servo stabilization gimbal is determined based on the current roll, pitch, and yaw angles. Specifically, the current hull attitude angles are read, and combined with the coordinate direction relationship determined during the installation of the servo stabilization gimbal, the hull roll direction is mapped to the gimbal roll compensation direction, the hull pitch direction to the gimbal pitch compensation direction, and the hull yaw direction to the gimbal heading compensation direction. For each compensation direction, the current hull attitude angle is converted into the gimbal target rotation angle in the opposite direction, and then a drive command is generated according to the angle command format of the gimbal controller. The servo stabilization gimbal rotates according to the drive command, making the attitude change direction of the shipborne 3D radar opposite to the attitude change direction of the hull, thus canceling the hull motion at the mechanical execution level.

[0043] After completing mechanical attitude compensation, the actual execution attitude of the servo-stabilized gimbal and the real-time attitude of the hull are read, and these two are used as the data basis for subsequent residual attitude correction. Thus, mechanical compensation is responsible for eliminating hull attitude changes during the point cloud acquisition stage, while coordinate correction is responsible for handling the attitude deviations that remain after mechanical execution. The two processes work together on the same batch of point cloud data.

[0044] In the above implementation, step three involves determining the dust interference level based on the point cloud quality indicators from the multi-source acquired data. First, a fixed search range is established centered on each point to be analyzed, and the number of neighboring points within the search range is counted. Then, the local point density is calculated based on the number of neighboring points and the spatial size of the search area. For each neighboring point, the spatial distance between it and the target point is calculated, and the dispersion of the neighboring distance is calculated based on all spatial distances. Simultaneously, the number of isolated points that cannot form a continuous spatial connection with surrounding points is counted, and the proportion of isolated points is determined based on the relationship between the number of isolated points and the total number of points within the analysis area.

[0045] For the degree of local height jump, the height coordinates of each point within the same spatial neighborhood are read, the maximum and minimum height values ​​are determined, and the maximum height value is subtracted from the minimum height value to obtain the height change of the neighborhood. For the spatial continuity of the neighborhood, it is determined whether adjacent points form a continuous connection according to the preset adjacency distance, and the number of points that meet the continuous connection condition is counted. The local point density, isolated point ratio, neighborhood distance dispersion, height change, and neighborhood spatial continuity are compared with the pre-established dust judgment thresholds. The comparison results of each index are used to characterize the spatial dispersion, surface continuity, and anomalous echo characteristics of the point to be analyzed or its local area, and serve as local quality features for subsequent point-level confidence assessment, dust outlier identification, and suspected boundary point retention.

[0046] It should be noted that the aforementioned local point cloud quality features are used for anomaly analysis of individual points or local point cloud regions, and are not directly used as the output basis for the dust interference level of the current metering cycle. The dust interference level of the current metering cycle is determined based on point density, reflection intensity fluctuation value, point cloud hole ratio, and multi-frame repeated observation rate. The former is used for local point-level dust anomaly identification, while the latter is used to characterize the degree of dust interference on the overall point cloud of the current metering cycle. The two are applied to different processing levels. Since there is no stable local surface connection between the echo points formed by suspended dust and the continuous stack surface, when the number of isolated points in the target area reaches the corresponding condition, the number of continuous connections in the neighborhood does not meet the corresponding condition, or the spatial distance between the target point and the surrounding stack surface reaches the outlier determination condition, the dust interference state of that area is classified into the corresponding level. After determining the dust interference level, the dust interference level, ship sway level, and current cargo type are used as parameter query conditions to read the corresponding point cloud acquisition parameters and point cloud processing parameters from the preset parameter mapping table. Each combination of conditions in the parameter mapping table corresponds to a defined scan parameter, time slice length, neighborhood search range, voxel size, outlier detection threshold, completion adjacency range, and surface reconstruction scale. Therefore, the current processing parameters are directly determined by the cargo type, ship sway level, and dust interference level.

[0047] In the above implementation, step four involves dividing the continuously acquired dry bulk cargo point cloud into multiple time slices based on the timestamps in the multi-source acquisition data. First, the length of the corresponding time slice is read according to the current hull sway level. Then, starting from the beginning of continuous acquisition, the start and end times of each time slice are determined sequentially according to its length. Point clouds acquired between the same start and end times are grouped into the same time slice, and the hull attitude data and servo stabilization gimbal attitude data within that time slice are associated with the corresponding point cloud.

[0048] For each time slice, the combined coordinate transformation relationship used for residual attitude correction is pre-calculated. Specifically, firstly, during the equipment installation and calibration phase, the position offset and installation angle of the shipborne 3D radar relative to the servo stabilization gimbal are determined and saved as the first set of fixed transformation parameters. Then, based on the actual roll, pitch, and yaw angles of the servo stabilization gimbal in the current time slice, the second set of transformation parameters of the servo stabilization gimbal relative to the ship is determined. Subsequently, based on the roll, pitch, and bow angles of the ship in the current time slice, the third set of transformation parameters of the ship relative to the reference coordinate system is determined. During the combined calculation, points in the 3D radar coordinate system are first transformed to the servo stabilization gimbal coordinate system according to the first set of fixed transformation parameters; then, the transformed points are transformed from the servo stabilization gimbal coordinate system to the ship coordinate system according to the second set of transformation parameters; finally, points in the ship coordinate system are transformed to the unified reference coordinate system according to the third set of transformation parameters. The rotation and translation relationships in the above three-level transformations are pre-combined in a fixed order to form the combined coordinate transformation matrix corresponding to the current time slice. When processing all points within the same time slice, the combined coordinate transformation matrix is ​​called directly, and the three-level transformation parameter combination is no longer executed repeatedly for each point.

[0049] In the above implementation, step five involves transforming the dry bulk cargo point clouds in each time slice to the same reference coordinate system. For each point in the current time slice, the three spatial coordinate values ​​of that point in the three-dimensional radar coordinate system are first read. Then, according to the pre-calculated combined coordinate transformation matrix for the current time slice, spatial rotation and translation are performed sequentially on that point. After rotation and translation, the three spatial coordinate values ​​of that point in the reference coordinate system are obtained. After completing the above processing for all points in the current time slice, the remaining time slices are processed sequentially until all point clouds at different acquisition times and different acquisition perspectives are represented using the same reference coordinate system, thereby obtaining the corrected point cloud. The reference coordinate system is determined during the equipment calibration stage, and its origin and the directions of the three coordinate axes remain unchanged throughout the entire metrology task. Through unified coordinate transformation, the point clouds corresponding to the same actual stack location at different acquisition times are transformed to the same spatial reference, so that subsequent processing is no longer directly affected by changes in the radar coordinate system caused by changes in the ship's attitude.

[0050] In the above implementation, step six involves dividing the calibration point cloud into multiple spatial grid cells. First, the starting and ending points of the grid division, as well as the grid size, are determined based on the spatial boundary of the dry bulk cargo metering area. Then, starting from the dividing point, the grid is continuously divided along two horizontal directions according to the fixed grid size, forming spatial grid cells covering the entire metering area. Each spatial grid cell corresponds to a defined horizontal and vertical range. For each point in the calibration point cloud, its corresponding spatial grid cell is determined based on its horizontal coordinates, and the point is written into the data set of the corresponding spatial grid cell. After all points are assigned, the number of points, point density, number of valid points, height dispersion, number of neighboring continuous points, number of outliers, and point cloud missing status are calculated for each spatial grid cell.

[0051] Point density is obtained by dividing the number of points within a grid by the corresponding spatial area or volume. Height dispersion is obtained by determining the maximum and minimum height coordinates within the grid and calculating the difference between them. The number of consecutive neighboring points is obtained by statistically analyzing points that meet preset adjacency distance conditions. The number of outliers is obtained by statistically analyzing points that do not meet neighborhood connectivity conditions or meet outlier criteria. These point cloud quality features are compared with corresponding grid quality thresholds, and the grid quality level of each spatial grid unit is determined according to preset grading rules.

[0052] In the above implementation, step seven involves dividing each spatial raster unit into a fine-processing region and a simplified-processing region based on the raster quality level. Spatial raster units that meet the first processing condition are assigned to the fine-processing region, while those that meet the second processing condition are assigned to the simplified-processing region. Both the first and second processing conditions are determined by a combination of thresholds for the number of points, point density, height dispersion, number of outliers, point cloud missing status, and boundary position status. For the fine-processing region, a point-level reliability assessment is performed on each point within the region. A fixed neighborhood search range corresponding to the current cargo type is established with the point to be evaluated as the center, the number of neighborhood points is counted, and a local reference plane is established based on the neighborhood points. Subsequently, the vertical distance from the point to be evaluated to the local reference plane is calculated, and the surface orientation difference, neighborhood height change, and coordinate deviation of the same spatial position in adjacent time slices are calculated between the neighborhood of the point to be evaluated and its adjacent neighborhoods.

[0053] The calculation results are compared with the corresponding thresholds. If a target point simultaneously meets the criteria for determining a real stack surface, it is classified into the real stack surface category. If a target point meets the criteria for determining dust outliers, it is classified into the dust category. If a target point is located at the stack surface outline, cargo hold boundary, or a location with continuous local height transitions, and meets the boundary retention criteria, it is classified into the boundary category. These three categories correspond to retention, deletion, and boundary retention processing, respectively.

[0054] For simplified processing areas, adaptive voxelization downsampling is performed. First, the base voxel size corresponding to the current cargo type is read. Then, based on the point density interval of the current spatial raster, the corresponding voxel size adjustment value is read from the parameter table to obtain the actual voxel size used by the spatial raster. The spatial raster is further divided into multiple voxels according to this size. For each voxel, the three coordinate values ​​of all valid points are counted. All coordinate values ​​in the same direction are added together and then divided by the number of valid points within the voxel to obtain the average coordinate value in the three directions. The position determined by the three average coordinate values ​​is used as the representative position of the voxel. Alternatively, the point with the smallest spatial distance from the average position can be selected from the actual collected points within the voxel as the representative point. After voxelization, only one representative point is retained for each non-empty voxel. After voxelization downsampling, lightweight filtering is performed. A neighborhood search range with a fixed radius is established centered on each representative point, and the number of valid points within the neighborhood is counted. When the number of valid points does not reach the threshold of the minimum number of neighboring points corresponding to the current cargo type, the representative point is marked as an isolated point and deleted; when the threshold is reached, the representative point is retained.

[0055] In the above implementation, step eight is performed to suppress dust outliers based on the point-level confidence assessment results and the adaptive metering parameters for the cargo type. For the current cargo type, the neighborhood search distance, local reference surface distance threshold, neighborhood connection number threshold, and boundary preservation conditions are read from the cargo type parameter group.

[0056] For each point to be judged, first count the number of its neighboring connected points, then calculate the distance from the point to the local reference surface in the neighborhood, and determine whether the point is located in a pre-identified stack surface boundary grid or cargo hold boundary adjacent grid. Points that meet the true stack surface judgment criteria are marked as true stack surface points and retained; points whose number of neighboring connected points does not reach the set number and whose distance to the local reference surface reaches the outlier judgment threshold are marked as dust outliers and deleted; points located at stack surface boundaries, cargo hold boundaries, or height-continuous turning points and meeting the boundary retention criteria are marked as suspected boundary points and retained. After completing the dust outlier suppression, the spatial grid is used to identify missing regions. The number of true stack surface points in each spatial grid is checked sequentially. When there are no points in a spatial grid that meet the true stack surface judgment criteria, and at least one spatial grid directly adjacent to this spatial grid contains a true stack surface point or a suspected boundary point, the spatial grid is marked as a point cloud missing grid. The interconnected point cloud missing grids are merged to obtain the point cloud missing region. After identifying the missing point cloud region, a predetermined number of adjacent grid cells are expanded outward from the boundary of that region to form a local completion range. The completion process is only performed within the missing point cloud region and the local completion range, without changing the actual stacked points in other regions.

[0057] For point cloud missing regions located inside the stack surface, first read the spatial coordinates of the real stack surface points and suspected boundary points around the missing region; then determine the local slope at the boundary of the missing region based on the height changes of each real stack surface point in adjacent directions; then generate a continuous completion surface based on the boundary height and local slope on both sides or in multiple directions of the missing region, so that the position of the completion surface at the edge of the missing region and the real stack surface remains continuous.

[0058] For missing point cloud regions adjacent to the cargo hold boundary, in addition to using surrounding real stack surface points, the fixed position of the cargo hold boundary in the reference coordinate system is also read. When generating completion points, each completion point is checked to see if it is within the measurement space defined by the cargo hold boundary; completion points outside the cargo hold boundary are not written into the completed stack surface point cloud. Finally, the retained real stack surface points, suspected boundary points, and completed points are merged to obtain the completed stack surface point cloud.

[0059] In the above implementation, step nine involves adaptive surface reconstruction based on the completed pile surface point cloud, adaptive metering parameters for the cargo type, and the degree of local pile surface undulation. First, the local pile surface undulation characteristics are calculated for each spatial grid cell. When calculating height undulation, the height coordinates of all valid points in the current spatial grid are read, the maximum and minimum height coordinates are determined, and the difference between them is calculated. When calculating local surface orientation changes, local surface orientations are established based on valid points in the current and adjacent spatial grids, and the angle between the two local surface orientations is calculated. When calculating local fitting deviation, a local reference surface is first established using valid points in the current spatial grid, then the distance from each valid point to the local reference surface is calculated, and all distances are accumulated and divided by the number of valid points to obtain the local fitting deviation of the spatial grid.

[0060] The aforementioned local surface undulation features are compared with a pre-set undulation level threshold to determine the local surface undulation level corresponding to the current spatial grid. Each local surface undulation level corresponds to a pre-set surface reconstruction scale, and each cargo type corresponds to an independent set of reconstruction scale parameters. Therefore, after determining the local surface undulation level of the current cargo type and the current spatial grid, the corresponding surface reconstruction scale is directly read from the parameter set. When reconstructing using the grid height surface, the height coordinates of the actual surface points, suspected boundary points, and completion points within each spatial grid are statistically analyzed, and the neighborhood range participating in the height calculation is determined according to the current reconstruction scale. The height coordinates of the effective points within the neighborhood are added together and then divided by the number of effective points participating in the calculation to obtain the reconstructed height corresponding to that spatial location. The reconstructed heights are connected according to the arrangement order of the spatial grid to form a continuous dry bulk cargo surface.

[0061] When using triangular mesh surfaces for reconstruction, connecting edges are established based on the spatial adjacency relationship between each effective point. Then, three adjacent points are connected to form triangular patches according to non-intersecting adjacency relationships. Triangular patches are continuously generated between adjacent spatial grids until the corresponding area of ​​the stacked point cloud is completed to form a continuous stacked surface.

[0062] After obtaining the surface of the dry bulk cargo stack, the volume of the dry bulk cargo is calculated based on the measurement datum and the boundary of the measurement area. When using the spatial grid accumulation method, for each spatial grid involved in the calculation, first read the reconstructed stack height of the grid, then read the datum height corresponding to the same horizontal position; subtract the datum height from the stack height to obtain the effective stack height corresponding to the spatial grid; then multiply the effective stack height by the horizontal projected area of ​​the spatial grid to obtain the local cargo volume corresponding to the spatial grid; finally, add up the local cargo volumes of all effective spatial grids within the measurement area to obtain the total volume of the dry bulk cargo. When using the triangular mesh method for volume calculation, the reconstructed dry bulk cargo stack surface, the side boundary of the measurement area, and the measurement datum are connected to form a closed space. Then, the corresponding local closed volume is calculated according to the spatial coordinates of each triangular facet, and all local closed volumes are accumulated to obtain the total volume of the dry bulk cargo.

[0063] Through the above processing, cargo type is used to determine adaptive measurement parameters, ship attitude data is used to determine the ship sway level and servo reverse attitude compensation amount, point cloud quality index is used to determine the dust interference level, time slice is used to establish residual attitude correction relationship, spatial grid quality is used to determine the fine processing area and simplified processing area, point-level confidence level is used to distinguish between real stack points, dust outliers and suspected boundary points, local physical constraint completion is used to restore missing point cloud areas, and adaptive surface reconstruction and spatial volume accumulation are used to obtain the final dry bulk cargo volume measurement result. Each processing step uses pre-set parameters, thresholds, mapping relationships and calculation steps to determine the processing result, so that the entire measurement process has a defined data input, processing path and output result.

[0064] Experimental platform and data acquisition scheme:

[0065] Experimental site and subjects: No. 2 cargo hold (hatch opening approximately 40m × 25m) of a 50,000-ton bulk carrier at a port, loaded with four types of cargo: iron ore powder (powdered, high dust), coal (mixed powder and lumps), crushed stone (lumps), and soybeans (granular); each type of cargo was loaded with approximately 8,000-12,000 tons.

[0066] Acquisition hardware:

[0067] Two shipborne 3D lidar units (mechanical rotating type, ranging accuracy ±2cm, point frequency 3.2×10⁻⁶) 5 (points / s), installed diagonally on the mast above the hatch coaming;

[0068] Two sets of servo-stabilized gimbals (attitude stabilization accuracy ±0.05°, response bandwidth 10Hz), with the radar mounted on the gimbal;

[0069] The ship's attitude reference unit (MRU, 100Hz) outputs roll, pitch, and bow angles, which are synchronously recorded with the radar point cloud and the actual compensated attitude of the gimbal according to a unified PTP time reference.

[0070] The embedded industrial computer (Intel i7-1185G7 / 32GB) performs all point cloud processing and volume calculation.

[0071] True value acquisition (feasible in practice): After each working condition measurement is completed, the operation is paused, and the surveyor uses a total station to measure the elevation of the pile surface in a 2m×2m grid. The elevation is then cross-checked with the ship draft measurement results to obtain the reference true value volume V0 (its own uncertainty is about 0.5%, which is one order of magnitude lower than the error of the measured method and can be used as a benchmark).

[0072] Dust test conditions were constructed as follows: four test environments were set up: no dust, light dust, medium dust, and heavy dust. The no-dust condition involved the container remaining stationary without any loading or unloading operations. The light, medium, and heavy dust conditions involved continuous operation of the unloader's grab bucket, supplemented by axial flow fans, to create environments with varying degrees of suspended dust. The terms "no dust," "light dust," "medium dust," and "heavy dust" are test condition names used for performance verification and do not represent the system output levels of dust interference levels in this invention.

[0073] Under various experimental conditions, this invention calculates the point density, reflection intensity fluctuation value, point cloud hole ratio, and multi-frame repeated observation rate of the current measurement cycle according to S3. Based on the first and second classification thresholds of the corresponding cargo type, each indicator is divided into Level 1, Level 2, or Level 3 interference levels. The highest level among the corresponding levels of each indicator is then taken as the dust interference level of the current measurement cycle.

[0074] During the experiment, the proportion of isolated points, the degree of dispersion of neighborhood distance, height jump and neighborhood continuity were statistically analyzed to characterize the spatial dispersion and continuity changes of local point clouds after being affected by dust echoes, and to verify the subsequent point-level credibility assessment and dust outlier identification effect. The dust interference level was not generated separately based on this set of local indicators.

[0075] The ship's rolling levels were constructed as follows: steady state at berth (Level 1, rolling within ±0.3°), swell condition (Level 2, rolling ±1.5°), and tugboat berthing disturbance simulation severe condition (Level 3, rolling ±3° or more), all recorded by MRU actual measurements.

[0076] Comparison method:

[0077] Method A (Fixed Parameter Method): Without enabling stabilization gimbal compensation, the same set of fixed filtering, completion, and reconstruction parameters are used for all types of goods and all operating conditions;

[0078] Method B (Mechanical stabilization only): Enable gimbal reverse compensation, but do not perform time-slice residual attitude correction at the back end, and use fixed processing parameters;

[0079] Method C (the method of this invention): Complete steps S1-S9.

[0080] Evaluation indicators: Volume measurement relative error = |V_measured - V_0| / V_0 × 100%; True pile surface point retention rate = Number of true pile surface points retained / Total number of true pile surface points (judged with a tolerance of ±5cm for the total station grid surface); Dust outlier removal rate = Number of dust points removed / Total number of calibrated dust points; Single cycle processing time = Wall clock time from point cloud input to volume output (average value taken after 10 repetitions).

[0081] Experiment 1: Comparison of Volumetric Measurement Accuracy under Shipboard Dynamic Conditions

[0082] Operating conditions: coal cargo, surge conditions (secondary sloshing), medium dust interference; the three methods were used to repeatedly measure the same stockpile in the same compartment 6 times each.

[0083] Table 1. Raw data of relative error in volumetric measurement using three methods (%)

[0084]

[0085] As shown in Table 1, Method A has the largest error and the largest fluctuation, indicating that the fixed parameters cannot stably reconstruct the stack surface when the hull sway is superimposed with dust noise. After introducing mechanical stabilization, the mean value of Method B decreased by 1.36 percentage points, proving that the reverse attitude compensation is effective, but the residual attitude error and fixed parameters still limit the accuracy. Method C superimposes residual correction and adaptive parameters of the time slice combined coordinate transformation matrix on the basis of mechanical compensation, and the mean value is reduced to 0.86%, and the standard deviation is only 29% of that of A, and the repeatability is significantly improved.

[0086] Figure 6 As shown, the column height represents the mean, and the error bar represents the standard deviation. The method of this invention reduces the error by approximately 74.9% compared to the fixed parameter method and by approximately 58.3% compared to the mechanical stabilization method alone, verifying the superimposed gain of dual attitude processing of mechanical compensation and coordinate correction and the adaptive parameter link.

[0087] Experiment 2: Dust Interference Suppression and Real Stack Surface Retention Experiment

[0088] Operating conditions: Iron ore powder cargo, steady state (level 1 sway) at the berth. The same batch of point clouds was processed using methods A and C under four different dust disturbance levels, and the point classification results were statistically analyzed.

[0089] Table 2. Retention rate of real pile surface dust and removal rate of dust outliers under different dust test conditions.

[0090]

[0091] As shown in Table 2, Method A had to tighten the fixed threshold to remove dust under heavy dust conditions, resulting in 65% of the real pile surface points being mistakenly deleted (retention rate of only 0.35%), the pile surface being over-smoothed, and the volume being systematically undersized. The method of this invention distinguishes between real pile surface points, dust outliers, and suspected boundary points based on point-level reliability assessment, and combines the dust suppression parameters of the cargo type for adaptive judgment. Under heavy dust conditions, the retention rate still reaches 0.82, while the dust removal rate is increased from 0.70 to 0.95, that is, accurate deletion and complete retention.

[0092] Figure 7 As shown, the two ends of the dumbbell represent method A (gray) and method C (green), respectively, and the length of the line connecting them represents the difference between the two methods. The heavier the dust, the greater the difference between the two methods (the retention rate difference reaches 0.47 under heavy dust), indicating that the dust classification and point-level reliability mechanism of this invention has the most significant gain in harsh dust environments.

[0093] Experiment 3: Verification of Adaptive Metering Parameters for Multiple Commodities

[0094] Operating conditions: heavy dust disturbance, steady state at the berth. The four types of cargo are measured using methods A and C respectively, and each is repeated 3 times and the average value is taken.

[0095] Table 3. Relative error in volume measurement of different types of goods under heavy dust interference (%)

[0096]

[0097] As shown in Table 3, Method A uses the same set of filtering and reconstruction parameters for different types of goods. Powdered goods (iron ore powder, coal) have higher errors due to residual dust, while blocky goods (crushed stone) also have considerable errors due to excessive smoothing of the contour. The method of this invention switches the adaptive measurement parameter set (filtering, dust suppression, angle of repose, surface roughness, mesh reconstruction parameters) according to the type of goods. The errors of the four types of goods are reduced to less than 1% and the levels are similar, indicating that the adaptive mechanism of goods eliminates the inherent defect that a set of parameters cannot take into account multiple types of goods.

[0098] like Figure 8 As shown, the error of the method of the present invention is stable in a narrow range of 0.74%–0.97% for all types of goods, while the error of the fixed parameter method fluctuates between 2.44% and 3.28% depending on the characteristics of the goods. This verifies the adaptability of the adaptive parameter for each type of goods to the differences in particle size, roughness, and reflection characteristics of the materials.

[0099] Experiment 4: Real-time Verification of Hierarchical Processing

[0100] Operating conditions: coal type, surge condition. By adjusting the radar scan duration, a point cloud sequence of 10,000-120,000 points / frame is obtained, and the processing time of a single cycle (time slice division → coordinate unification → hierarchical processing → completion → reconstruction → volume output) is statistically analyzed.

[0101] Table 4. Processing time per cycle varies with point cloud size (s)

[0102]

[0103] As shown in Table 4, the time required for full-area fine processing increases approximately linearly with the number of points, reaching 6.08s for 120,000 points, which is difficult to meet the continuous metering cycle. The method of this invention concentrates the high-complexity processing on low-quality grids (dust area, missing area, material pile edge), and only performs adaptive voxelization downsampling and lightweight filtering on high-quality grids. The time required for 120,000 points is 2.56s, a reduction of 57.9%, and the growth slope is significantly slower, which verifies the compression effect of grid quality classification on computational redundancy.

[0104] like Figure 9 As shown, the distance between the two curves increases with the number of points, indicating that the larger the point cloud scale, the more prominent the real-time advantage of hierarchical processing, supporting continuous online measurement on ships.

[0105] Under shipboard dynamic and medium dust conditions, the volumetric measurement relative error of the method of this invention is 0.86%±0.11%, which is about 74.9% lower than the fixed parameter method (3.42%±0.38%) and about 58.3% lower than the mechanical stabilization method alone (2.06%±0.27%). It also exhibits the best repeatability, verifying the combined contributions of servo reverse compensation, time-slice residual attitude correction, and adaptive parameters. Under heavy dust conditions, the actual pile surface point retention rate is 0.82 (0.35 with the fixed parameter method), and the dust outlier removal rate is 0.95 (0.70 with the fixed parameter method). Point-level reliability assessment and cargo-specific dust parameters achieve a balance between dust suppression and profile preservation. The errors for four typical cargo types are all controlled within 1%, and the cargo-specific adaptive parameter set eliminates systematic deviations caused by differences in material characteristics. Graded processing reduces the single-cycle time of 120,000 points / frame from 6.08s to 2.56s (−57.9%), meeting the continuous measurement cycle requirements. The four experiments verified the beneficial effects described in the specification of this invention from four dimensions: dynamic accuracy, dust suppression, multi-cargo type adaptation and real-time performance. The data were all obtained based on a shipborne radar + stabilization gimbal + MRU platform and total station / water gauge true value system that can be practically built. The experimental scale was moderate and the steps were reproducible.

[0106] like Figure 5 As shown in the figures above and below, the comparison between the effects of cargo stack modeling and measurement processing based on 3D radar point clouds is presented in the dry bulk cargo measurement scenario.

[0107] (a) represents the original point cloud results from existing technologies or without sufficient optimization. Due to issues such as ship swaying, dust obstruction, multi-view acquisition errors, uneven point cloud density, local missing points, and outlier interference at dry bulk cargo loading and unloading sites, the original point cloud often exhibits obvious trailing shadows, ghosting, ripples, noisy scattered points, and missing areas on the stack surface. If such point clouds are directly used for surface reconstruction and volume calculation, it can easily lead to inaccurate stack boundary identification, deviations in stack height estimation, and increased volume integration errors, thereby affecting the accuracy and stability of dry bulk cargo weight measurement results.

[0108] (b) Completed pile surface point cloud or reconstructed model after processing using the method of the present invention. Through the processing flow proposed in this invention, the original messy, discontinuous point cloud, which is greatly affected by dust and shaking, can be transformed into a three-dimensional cargo pile model with clear boundaries, continuous pile surface, complete structure, and suitable for measurement calculation.

[0109] Example 2:

[0110] The multi-source data collected in S1 includes dry bulk cargo point cloud, radar acquisition timestamp, radar installation attitude data, servo stabilization gimbal compensation data, ship attitude data, shore-based reference data, and cargo type information.

[0111] Based on the cargo type information, a cargo type adaptive measurement parameter set is generated. The cargo type adaptive measurement parameter set includes point cloud filtering parameters, dust suppression parameters, point cloud completion parameters, stacking angle of repose parameters, surface roughness feature parameters, and mesh reconstruction parameters. The cargo type adaptive measurement parameter set is used to adapt parameters for point cloud processing, dust discrimination, boundary preservation, local completion, and surface reconstruction processes corresponding to different cargo types.

[0112] In the above implementation, S1 collects dry bulk cargo point cloud, radar acquisition timestamp, radar installation attitude data, servo stabilization gimbal compensation data, ship attitude data, shore-based reference data, and cargo type information. The shipborne 3D radar continuously collects dry bulk cargo stack echo data according to a set scanning cycle, and converts each echo point into a spatial coordinate point, forming a dry bulk cargo point cloud arranged in the acquisition sequence. After each frame of point cloud acquisition is completed, the acquisition time corresponding to the radar's internal clock is read and written into the data header of the current frame of point cloud to form a radar acquisition timestamp. Each spatial coordinate point in the point cloud is thus established with a definite association with the corresponding acquisition time, and subsequent ship attitude compensation and coordinate correction use the radar acquisition timestamp as a time index.

[0113] The radar installation pose data is obtained through calibration after the shipborne 3D radar is fixedly installed. During calibration, the rotation center of the servo-stabilized gimbal is used as the installation reference position. The distances of the 3D radar measurement origin relative to this reference position in three mutually perpendicular directions are measured, and the installation rotation angles of the three measurement coordinate axes of the 3D radar relative to the gimbal's reference coordinate axes are determined. The position offsets in the three directions and the installation rotation angles in the three directions are saved as radar installation pose data. Under the same installation conditions, this data remains unchanged and is directly called upon during each point cloud coordinate transformation, so that the points measured by the 3D radar are first transformed to the coordinate relationship corresponding to the servo-stabilized gimbal.

[0114] The servo-stabilized gimbal compensation data uses gimbal position feedback data, not target attitude commands issued by the controller. After driving the roll, pitch, and yaw axes to rotate, the gimbal controller reads the actual rotation angle fed back by the encoders of each axis and records the feedback time simultaneously. During point cloud processing, the radar acquisition timestamp of the current frame point cloud is used as the retrieval time. The preceding and following feedback data adjacent to this retrieval time are read from the gimbal compensation data. When the feedback time is consistent with the radar acquisition time, the feedback data is directly used; when the feedback time is inconsistent with the radar acquisition time, the time interval between the radar acquisition time and the preceding feedback time is first calculated, then the total time interval between the two feedback data is calculated. The two time intervals are divided to obtain the time-position coefficient. Then, the difference between the following rotation angle value and the preceding rotation angle value is calculated. This difference is multiplied by the time-position coefficient and added to the preceding rotation angle value. The result is used as the actual gimbal rotation angle corresponding to the radar acquisition time. The roll axis, pitch axis, and yaw axis are all processed according to the above process to obtain the servo stabilization gimbal compensation data corresponding to the current frame point cloud.

[0115] In this embodiment, the ship's attitude data includes roll, pitch, and bow angles, and is recorded according to the same time reference as the radar acquisition data. For cases where the sampling time of the ship's attitude sensor differs from the radar acquisition time, the same time correspondence processing method as the servo stabilization gimbal compensation data is used to determine the roll, pitch, and bow angles corresponding to the current radar acquisition time. Therefore, each frame of point cloud corresponds to one set of determined ship attitude data and one set of determined servo stabilization gimbal compensation data, providing a data basis at the same time for subsequent ship sway level determination and residual attitude correction.

[0116] The shore-based reference data is obtained from a reference device set at a fixed location on the shore. Before the measurement operation begins, the fixed spatial position of the shore-based reference point within the measurement area is determined, and the direction of the shore-based reference coordinate axes relative to the measurement area is also determined. After the ship enters the measurement area, the positional changes of the ship's reference position relative to the shore-based reference point in three directions are continuously acquired, and the ship's attitude data is recorded simultaneously. When processing the point cloud of the current frame, the measurement reference position is first determined based on the fixed position of the shore-based reference point. Then, the ship's translation is determined based on the distances in three directions between the ship's reference position and the shore-based reference point. Subsequently, the ship's attitude change relative to the shore-based reference coordinates is determined by combining the ship's roll, pitch, and bow angles. The obtained positional and attitude changes are used to subsequently convert the point clouds from each acquisition moment to the same reference coordinate system.

[0117] Cargo type information is recorded using pre-established cargo type numbers. Each type of dry bulk cargo corresponds to a unique cargo type number. Before executing the metering task, the current cargo type number is read from the loading and unloading task data, and the cargo type parameter table is retrieved using that number. The cargo type parameter table stores point cloud filtering parameters, dust suppression parameters, point cloud completion parameters, angle of repose parameters, surface roughness feature parameters, and mesh reconstruction parameters according to the cargo type number. After retrieving the current cargo type number, all parameters corresponding to that number are read into the parameter cache of the current metering task, forming the cargo type adaptive metering parameter set. Other cargo type parameters are not reused in subsequent processing until the cargo type number changes.

[0118] In this embodiment, the point cloud filtering parameters are used to limit the neighborhood search distance, neighborhood point number threshold, isolated point distance threshold, and voxel size for the current cargo type. During point cloud filtering, a spherical neighborhood is established centered on the point to be judged, and the radius of the spherical neighborhood is directly read from the neighborhood search distance corresponding to the current cargo type. Then, the coordinate differences between other points within the neighborhood and the point to be judged in the three coordinate directions are calculated point by point. The three coordinate differences are squared and summed, and the square root of the sum is taken to obtain the spatial distance between the two points. Points whose spatial distance does not exceed the neighborhood search distance are counted in the neighborhood point count. After the statistics are completed, the neighborhood point count is compared with the neighborhood point number threshold for the current cargo type. Points that reach the threshold proceed to subsequent processing, while points that do not reach the threshold are treated as isolated points. When performing voxelization downsampling, the point cloud space is divided into equal-sized sections along three coordinate directions based on the voxel size corresponding to the current cargo type. For any spatial point, the differences between its three coordinate values ​​and the raster's starting coordinates are calculated, and then divided by the voxel size. The integer part of the quotient is taken as the voxel number of that point in the three directions. Points with the same three voxel numbers are grouped into the same voxel. For each non-empty voxel, the coordinate values ​​of all points within the voxel in the three coordinate directions are accumulated, and then divided by the number of points within the voxel to obtain the three average coordinate values ​​of the voxel. The original point with the smallest spatial distance from this average position is retained as the representative point, and the remaining points are deleted. After this processing, only one true sampling point is retained within the voxel, avoiding duplicate sampling points from entering subsequent calculations.

[0119] The dust suppression parameter is used to define the neighborhood connectivity and point-to-surface distance conditions for dust outliers. For the filtered target point, neighboring points are first searched based on the dust type corresponding to the current product, and the number of points forming a valid adjacency with the target point is counted. Then, a local reference surface is established using the points within the neighborhood that meet the valid conditions. When establishing the local reference surface, the average coordinates of all valid points within the neighborhood in three coordinate directions are calculated, and this average position is used as the local reference position. The local surface direction is then determined based on the spatial distribution of neighborhood points relative to this reference position. After the local reference surface is determined, the perpendicular distance from the target point along the local surface normal direction to the reference surface is calculated.

[0120] In this embodiment, when the number of effective adjacent points of a target point does not reach the connection number threshold for the current cargo type, and the vertical distance from the target point to the local reference surface reaches the dust distance threshold for the current cargo type, the point is identified as a dust outlier and deleted from the set of points to be reconstructed. When the number of effective adjacent points reaches the connection number threshold, and the vertical distance from the point to the local reference surface does not reach the dust distance threshold, the point is retained as a candidate point for the actual pile surface. By simultaneously using the neighborhood connection state and the local point-to-surface distance for judgment, the deletion of actual undulation points on the particle surface is avoided by relying solely on a single distance condition. The surface roughness feature parameters are obtained from the actual pile surface calibration data of the current cargo type. When establishing the parameters, a continuous point cloud region is selected in the calibration pile surface without dust interference. Local reference surfaces are established point by point according to the specified neighborhood size, and the vertical distance from each point in the neighborhood to the corresponding local reference surface is calculated. Then, the absolute values ​​of all vertical distances are summed and divided by the number of points involved in the calculation to obtain the average surface deviation of the calibration area; at the same time, the difference between the maximum height coordinate and the minimum height coordinate in the neighborhood is calculated to obtain the local height fluctuation. The average surface deviation range and local height fluctuation range determined by the product type calibration are saved as the surface roughness characteristic parameters of that product type.

[0121] In actual point cloud processing, the same calculations are performed on the neighborhood of the candidate points on the actual stack surface. When the average surface deviation and local height fluctuation of the neighborhood of the target point both fall within the range of the surface roughness feature parameters corresponding to the current cargo type, the local fluctuation is identified as a stack surface feature formed by the particle structure of the current cargo type itself, and is not deleted as a dust anomaly. When the target point is located in the stack surface outline grid or the adjacent grid of the cargo hold boundary, and its local surface features meet the surface roughness limit of the current cargo type, the target point is marked as a suspected boundary point and retained in subsequent point cloud completion and surface reconstruction.

[0122] Point cloud completion parameters include the number of adjacent grid cells to be completed, the completion search distance, and the completion height continuity threshold. After identifying missing point cloud regions, the number of adjacent grid cells specified for the current cargo type is expanded outward from the boundary of the missing region, and the actual stack points and suspected boundary points in the expanded region are read. For each location to be completed within the missing region, the nearest valid stack point is found along the four directions of front, back, left, and right, and the height and horizontal distance of the valid points in the four directions are recorded. Then, the change in height per unit horizontal distance between the valid points in each direction and the location to be completed is calculated, and the weights are determined according to the horizontal distance of the valid points in each direction to the location to be completed, with larger weights for directions that are closer. The height of the valid points in each direction is multiplied by their corresponding weights, summed, and then divided by the sum of all weights. The resulting height is used as the initial completion height of the location to be completed.

[0123] After obtaining the initial completion height, this height is compared with the heights of the four adjacent valid points in each of the four directions. The height difference is calculated, and then the height difference is divided by the corresponding horizontal distance to obtain the slope change value between the position to be completed and the adjacent valid points. If the slope change value does not reach the completion height continuity threshold corresponding to the current cargo type, the completion point is retained; if it reaches the completion height continuity threshold, the completion height is recalculated according to the remaining valid directions. If none of the valid directions meet the completion height continuity condition, the completion point at that position is not generated to prevent the formation of an unfounded completion surface at a location lacking effective geometric constraints.

[0124] The stacking repose angle parameter is used to limit the local slope of the completed and reconstructed surfaces. For two adjacent stacking locations, the height difference between the two locations is first calculated, and then the actual distance between the two locations on the horizontal plane is calculated. The local slope is determined by the ratio of the absolute value of the height difference to the horizontal distance, and then converted into the corresponding slope angle. This slope angle is compared with the stacking repose angle parameter for the current cargo type. When the slope angle does not reach the corresponding stacking repose angle, the current completed point or mesh connection is retained; when the slope angle reaches the corresponding stacking repose angle, the increase of the height difference along this direction is limited, and the completed height is re-determined based on the adjacent actual stacking points. This ensures that the local slope after point cloud completion and stacking reconstruction meets the stacking characteristics of the current cargo type.

[0125] In this embodiment, the grid reconstruction parameters include grid size, neighborhood connection distance, and local surface reconstruction scale. After obtaining the actual stacking points, suspected boundary points, and completion points, the measurement area is horizontally gridded according to the grid size of the current cargo type. For each grid, valid stacking points within its boundary range are collected. When there are more than two valid stacking points in a grid, the height values ​​of each valid point are summed and divided by the number of valid points to obtain the basic stacking height of that grid. Subsequently, adjacent grids are searched using the local surface reconstruction scale corresponding to the current cargo type, and their slope angles are checked based on the height difference and horizontal distance between adjacent grids. Adjacent grids that meet the stacking repose angle constraint of the current cargo type are connected by a continuous surface. Grids that do not meet the constraint have their local heights re-determined based on the retained actual stacking points and suspected boundary points.

[0126] Through the above processing, the cargo type information first determines a unique set of adaptive measurement parameters for the cargo type; point cloud filtering parameters are used to complete neighborhood screening and voxel downsampling; dust suppression parameters are used to identify dust outliers based on the number of neighborhood connections and point-to-surface distance; surface roughness feature parameters are used to distinguish between real surface undulations and outliers formed by cargo particles, and participate in the retention of suspected boundary points; point cloud completion parameters determine the completion range and height of missing regions; stacking angle of repose parameters limit the local slope after completion and reconstruction; and mesh reconstruction parameters determine the spatial discrete scale and connection range of the stack surface. All parameters are directly indexed from the current cargo type information and executed according to the determined calculation steps in the corresponding processing stages, thus ensuring that point cloud processing, dust discrimination, boundary preservation, local completion, and surface reconstruction for different cargo types have clear data sources, calculation processes, and judgment conditions.

[0127] Example 3:

[0128] In this embodiment, S2 specifically includes:

[0129] Extract hull attitude data and radar acquisition timestamps from multi-source data, and synchronize the hull attitude data according to the radar acquisition timestamps.

[0130] The amplitude of the hull attitude change is calculated based on the synchronized hull attitude data, and the hull rolling level is determined based on the amplitude of the hull attitude change.

[0131] The target reverse compensation attitude is determined based on the synchronized hull attitude data, and the compensation gain, compensation rate and compensation limit parameters of the servo stabilization gimbal are determined based on the hull sway level.

[0132] Based on the target's reverse compensation attitude and the compensation gain, compensation rate, and compensation limiting parameters, a gimbal compensation command is generated, causing the servo stabilization gimbal to drive the shipborne three-dimensional radar to perform attitude compensation opposite to the direction of the ship's sway.

[0133] Obtain the actual compensated attitude of the servo-stabilized gimbal and write the hull sway level and actual compensated attitude into the timing compensation data.

[0134] In the above implementation, hull attitude data and radar acquisition timestamps are first read from multi-source acquisition data. The hull attitude data includes roll, pitch, and bow angles, each with the sampling time of the attitude sensor. The radar acquisition timestamp corresponds to the acquisition time of each frame of dry bulk cargo point cloud. Using the radar acquisition timestamp of each frame of point cloud as a synchronization reference, the hull attitude data is searched for the first set of attitude data with the smallest time difference before the radar acquisition time, and the second set of attitude data with the smallest time difference after the radar acquisition time.

[0135] When the radar acquisition time coincides with the sampling time of one set of attitude data, that set of attitude data is directly used as the synchronized attitude data corresponding to the current frame point cloud. When the radar acquisition time is between the sampling times of two sets of attitude data, the sampling time of the previous set of attitude data is first subtracted from the radar acquisition time to obtain the first time interval, and then the sampling time of the previous set of attitude data is subtracted from the sampling time of the subsequent set of attitude data to obtain the second time interval. The first time interval is divided by the second time interval to obtain the time position ratio between the current radar acquisition time and the two attitude sampling times. Subsequently, the roll angle, pitch angle, and yaw angle are synchronized respectively. That is, the change in attitude angle is obtained by subtracting the previous set of attitude angles from the subsequent set of attitude angles, and then the change in attitude angle is multiplied by the time position ratio. The result is added to the previous set of attitude angles to obtain the synchronized roll angle, synchronized pitch angle, and synchronized yaw angle corresponding to the current radar acquisition time.

[0136] After time synchronization is completed, the amplitude of hull attitude changes is calculated according to continuous radar acquisition times. For the roll direction, the synchronous roll angle at the current radar acquisition time is subtracted from the synchronous roll angle at the previous radar acquisition time, and the absolute value of the difference is taken to obtain the roll change amplitude. The pitch change amplitude and bow change amplitude are calculated in the same way. Then, the roll change amplitude, pitch change amplitude, and bow change amplitude are compared with the pre-calibrated two-level threshold. When the attitude change amplitude in all three directions is less than the corresponding first threshold, it is determined to be level 1 hull sway; when the attitude change amplitude in any direction reaches the corresponding first threshold but is less than the corresponding second threshold, it is determined to be level 2 hull sway; when the attitude change amplitude in any direction reaches the corresponding second threshold, it is determined to be level 3 hull sway. Roll, pitch, and yaw are each assigned a first and second threshold value, respectively. These threshold values ​​are determined by calibrating the permissible attitude deviation of the shipborne 3D radar and the response performance of the servo stabilization gimbal, and then written into the parameter table. After the hull sway level is determined, the target's reverse compensation attitude is calculated based on the synchronized hull attitude data. First, the current synchronized roll, pitch, and yaw angles are read. Then, the directional correspondence between the hull coordinate axes and the roll, pitch, and yaw axes of the servo stabilization gimbal is determined based on the radar's installation attitude data. When the ship's coordinate axes are aligned with the directions of each axis of the gimbal, the synchronous roll angle, synchronous pitch angle, and synchronous yaw angle are changed in positive and negative directions respectively, serving as the basic reverse compensation angles for the roll axis, pitch axis, and yaw axis. When there are installation angles, the ship's attitude direction is first rotated in the coordinate system according to the three directional angles obtained from the installation calibration, and the ship's attitude is converted to the gimbal coordinate system. Then, the positive and negative directions of the three attitude angles after conversion are changed to obtain the basic reverse compensation angles for the three axes of the gimbal.

[0137] The compensation gain, compensation rate, and compensation limit parameters of the servo stabilization gimbal are read from the parameter table according to the hull roll level. The first, second, and third hull roll levels each correspond to a set of determined compensation gains, maximum compensation rates, and positive and negative limit angles. The compensation gain is used to correct the basic reverse compensation angle. Specifically, the basic reverse compensation angles of the roll, pitch, and yaw axes are multiplied by the compensation gain corresponding to the current hull roll level to obtain the gain compensation angles for the three axes.

[0138] In this embodiment, after the gain compensation processing is completed, a compensation rate limit is applied to the gain compensation angle. For any axis, the gain compensation angle of the current control cycle is subtracted from the target compensation angle issued in the previous control cycle to obtain the angle change required for the current control cycle. Then, the maximum compensation rate corresponding to the current hull roll level is multiplied by the duration of a single control cycle to obtain the maximum compensation angle that can be changed in a single control cycle. When the absolute value of the required angle change is less than the maximum compensation angle, the current gain compensation angle is directly used. When the absolute value of the required angle change reaches the maximum compensation angle, the maximum compensation angle is increased or decreased according to the current target change direction, starting from the target compensation angle of the previous control cycle, to obtain the rate-limited target compensation angle. The above processing is performed on the roll axis, pitch axis, and yaw axis respectively. Subsequently, compensation amplitude is applied to the target compensation angles of the three axes. The positive and negative limit angles of each axis are read respectively, and the rate-limited target compensation angle is compared with the corresponding limit angle. The target compensation angle is maintained when it is between the positive and negative limiting angles; when the target compensation angle reaches the positive limiting angle, the final target compensation angle is limited to the positive limiting angle; when the target compensation angle reaches the negative limiting angle, the final target compensation angle is limited to the negative limiting angle. After compensation gain processing, compensation rate limiting, and compensation amplitude limiting, the final target reverse compensation attitudes for the roll axis, pitch axis, and yaw axis are formed.

[0139] The process involves generating gimbal compensation commands based on the final target's reverse attitude compensation. Each gimbal compensation command includes the target angles for the roll, pitch, and yaw axes, as well as the target compensation rates for the three axes. Upon receiving the compensation commands, the gimbal controller reads the current actual rotation angles fed back from the encoders of the three axes and subtracts the actual rotation angles from the target angles to obtain the angle deviations of the corresponding axes. The sign of the angle deviation determines the rotation direction of the servo motors, and the rotation speed is determined based on the compensation rate corresponding to the current hull roll level. This drives the roll, pitch, and yaw axes to rotate towards the target angles. Consequently, the roll axis rotates in the opposite direction to the hull roll, the pitch axis rotates in the opposite direction to the hull pitch, and the yaw axis rotates in the opposite direction to the hull bow roll, ensuring that the attitude change direction of the shipborne 3D radar is opposite to the hull attitude change direction.

[0140] Specifically, during the gimbal compensation process, the controller reads the encoder feedback values ​​of each axis according to a fixed control cycle and recalculates the angular deviation between the target angle and the actual rotation angle. If the angular deviation does not fall within the positioning error range determined by the equipment calibration, the corresponding axis continues to be driven at the current compensation rate. Once the angular deviation enters the positioning error range, the axis stops increasing the compensation angle and maintains its current execution position until a new target reverse compensation attitude is generated in the next control cycle. The actual compensation attitude is determined by the servo-stabilized gimbal encoder feedback value. For each frame of dry bulk cargo point cloud, the gimbal feedback data is retrieved according to the radar acquisition timestamp of that frame. When the feedback time coincides with the radar acquisition time, the corresponding actual roll axis angle, pitch axis angle, and yaw axis angle are directly read. When the feedback time does not coincide with the radar acquisition time, the two sets of gimbal feedback data that are closest to each other before and after the radar acquisition time are selected respectively. Following the aforementioned time synchronization calculation process, the time position ratio between the radar acquisition time and the two feedback times is first determined. Then, the actual compensated attitude corresponding to the radar acquisition time is calculated based on the difference between the two sets of roll angle, pitch angle, and yaw angle.

[0141] After obtaining the actual compensated attitude, a time-series compensation data record is established using the radar acquisition timestamp. Each record is written with the same data sequence: radar acquisition timestamp, hull roll level, actual roll axis compensation angle, actual pitch axis compensation angle, and actual yaw axis compensation angle. An index relationship is established with the dry bulk cargo point cloud corresponding to that radar acquisition timestamp. When S4 executes time slice division, the corresponding time-series compensation data is read according to the radar acquisition timestamp of each point cloud, and the actual compensated attitude is used to participate in residual attitude correction, so that subsequent coordinate transformations adopt the real attitude already formed by the gimbal, rather than the target compensated attitude generated by the controller.

[0142] Example 4:

[0143] In this embodiment, S3 specifically includes:

[0144] Dry bulk cargo point clouds are extracted from multi-source data, and the point density, reflection intensity fluctuation, point cloud hole ratio, and multi-frame repeated observation rate of the dry bulk cargo point clouds are calculated.

[0145] The dust interference level is determined based on the point cloud quality indicators.

[0146] Match the dust disturbance level with the hull sway level to determine the current metering conditions;

[0147] Based on the current metering conditions and the point cloud filtering parameters and dust suppression parameters in the adaptive metering parameter set for the cargo type, adjust the point cloud acquisition parameters and processing parameters.

[0148] Write the adjusted point cloud acquisition and processing parameters into the current metering cycle so that subsequent time slice division, coordinate unification, raster quality grading, and point cloud processing are performed according to the adjusted parameters.

[0149] In the above implementation, S3 uses the dry bulk cargo point cloud within the current metering cycle as the quality evaluation object. First, points outside the cargo hold metering range are deleted according to the metering area boundary. Then, the metering area is divided into several detection grids according to a preset horizontal grid size. Point cloud quality indicators are calculated within the same grid size and the same metering cycle to avoid incomparability between metering cycles due to changes in the calculation range. Point density is calculated based on the number of valid points within the detection grid and the horizontal area of ​​the detection grid. For each detection grid, the number of valid points falling into that grid is counted, and the point density of that detection grid is obtained by dividing the number of valid points by the horizontal area of ​​the grid. After completing the calculation of all detection grids, the point densities of each detection grid are added together and then divided by the number of detection grids involved in the calculation. The result is used as the point density for the current metering cycle. Non-stacking areas already determined during cargo hold structure calibration are not included in the point density calculation.

[0150] In this embodiment, the reflection intensity fluctuation value is calculated using the radar reflection intensity of valid points within the current metering period. For each detection grid, the reflection intensity values ​​of all valid points within the grid are first added together, then divided by the number of valid points in that grid to obtain the average reflection intensity. Subsequently, the difference between the reflection intensity of each valid point and the average reflection intensity is calculated, and the absolute value is taken. The absolute values ​​are then summed and divided by the number of valid points to obtain the reflection intensity fluctuation value of that grid. The reflection intensity fluctuation values ​​of all valid detection grids are added together and divided by the number of valid detection grids to obtain the reflection intensity fluctuation value for the current metering period. Using the above calculation method, the stability of the echo from the continuous stack surface of the same type of cargo can be converted into a definite value, and the reflection dispersion caused by dust echoes can be identified accordingly.

[0151] The point cloud hole ratio is determined based on the effective occupancy of the detection grid. For each detection grid, the number of valid points within the grid is counted. When the number of valid points reaches the minimum valid point threshold corresponding to the current cargo type, the grid is recorded as a valid grid; otherwise, it is recorded as a hole grid. The total number of hole grids within the current measurement period is counted, and then divided by the total number of detection grids participating in the point cloud quality evaluation. The result is used as the point cloud hole ratio. Fixed obstruction areas that cannot be observed by 3D radar during the equipment installation and calibration phase are pre-removed from the detection area and are not included in the number of hole grids or the total number of detection grids.

[0152] Specifically, the multi-frame repetition rate is calculated using the spatial correspondence between consecutive point cloud frames. First, the current frame and its preceding frame's point cloud are selected. Based on the radar acquisition timestamps corresponding to the two point clouds, the ship's attitude data and the actual compensated attitude of the servo-stabilized gimbal are read. The preceding frame's point cloud is then converted to the radar observation position corresponding to the acquisition time of the current frame to eliminate the influence of recorded attitude changes between the two acquisitions on the repetition judgment. Subsequently, each valid point in the current frame is used as a matching point, and the point with the smallest spatial distance in the preceding frame's point cloud is searched. The spatial distance is obtained by calculating the coordinate differences between the two points in three coordinate directions, squaring each of the three coordinate differences, adding them together, and then taking the square root of the sum.

[0153] Specifically, when the spatial distance does not reach the repeat observation distance threshold corresponding to the current cargo type, the point to be matched is recorded as a repeat observation point; when the spatial distance reaches the repeat observation distance threshold, the point is not counted as a repeat observation point. The repeat observation rate of the current frame is obtained by dividing the number of repeat observation points in the current frame by the number of valid points in the current frame. The above calculation is performed sequentially on adjacent point cloud frames within the current measurement period, and the repeat observation rates of each frame are added together and divided by the number of frames involved in the calculation to obtain the multi-frame repeat observation rate of the current measurement period. The actual stack surface maintains a definite spatial continuity at adjacent acquisition times, while the position of suspended dust changes with airflow and loading / unloading processes. Therefore, the multi-frame repeat observation rate can be used to characterize the temporal stability of echo points.

[0154] Specifically, after obtaining the point density, reflection intensity fluctuation value, point cloud hole ratio, and multi-frame repeat observation rate, these four indicators are compared with the dust classification thresholds corresponding to the current cargo type. The cargo type adaptive metering parameter set pre-stores the first point density threshold, the second point density threshold, the first reflection intensity fluctuation threshold, the second reflection intensity fluctuation threshold, the first hole ratio threshold, the second hole ratio threshold, the first repeat observation rate threshold, and the second repeat observation rate threshold for each cargo type.

[0155] Each point cloud quality index is divided into three interference levels by a first grading threshold and a second grading threshold. The three interference levels correspond to level 1 dust interference, level 2 dust interference, and level 3 dust interference, respectively.

[0156] For point density and multi-frame repeat observation rate, lower values ​​indicate a higher degree of dust interference in the point cloud. Specifically, when the point density reaches the second point density threshold, the corresponding interference level is determined as Level 1; when the point density is less than the second point density threshold but reaches the first point density threshold, it is determined as Level 2; and when the point density is less than the first point density threshold, it is determined as Level 3. The multi-frame repeat observation rate follows the same classification direction: when the multi-frame repeat observation rate reaches the second repeat observation rate threshold, it is determined as Level 1; when it is less than the second repeat observation rate threshold but reaches the first repeat observation rate threshold, it is determined as Level 2; and when it is less than the first repeat observation rate threshold, it is determined as Level 3.

[0157] For reflection intensity fluctuation and point cloud hole ratio, higher index values ​​indicate a higher degree of dust interference in the point cloud. Specifically, a reflection intensity fluctuation value less than the first reflection intensity fluctuation threshold is classified as Level 1, reaching the first reflection intensity fluctuation threshold but less than the second reflection intensity fluctuation threshold is classified as Level 2, and reaching the second reflection intensity fluctuation threshold is classified as Level 3. The point cloud hole ratio is judged according to the same classification direction: a point cloud hole ratio less than the first hole ratio threshold is classified as Level 1, reaching the first hole ratio threshold but less than the second hole ratio threshold is classified as Level 2, and reaching the second hole ratio threshold is classified as Level 3.

[0158] The grading direction of the four indicators is determined separately when calibrating the cargo parameters, and the judgment is not made by directly adding the four original values. After the grading of the four indicators is completed, the interference level with the largest level number is taken as the current dust interference level. Specifically, when all four indicators correspond to level 1, the dust interference level is determined to be level 1; when at least one indicator corresponds to level 2 and there is no level 3 indicator, the dust interference level is determined to be level 2; when at least one indicator corresponds to level 3, the dust interference level is determined to be level 3. Thus, any one of the point cloud holes, reflection anomalies, spatial sparsity, and temporal repeatability anomalies that meets the corresponding grading conditions can be reflected in the final dust interference level. Therefore, in this embodiment, the system output result of the dust interference level is three levels: level 1, level 2, and level 3. Two grading thresholds are used to divide each point cloud quality indicator into three continuous intervals, and there is no situation where two thresholds are used to directly distinguish four dust interference levels.

[0159] Specifically, after the dust interference level is determined, the ship sway level in the timing compensation data written by S2 is read, and the corresponding level is selected according to the radar acquisition timestamps covered by the current measurement cycle. When there are multiple ship sway levels in the current measurement cycle, the ship sway level with the largest level number in that measurement cycle is used as the ship sway level for the current cycle, so as to avoid attitude changes that have already occurred in the current cycle being covered by subsequent stable periods.

[0160] Dust disturbance level and hull sway level are matched using a pre-stored operating condition matching table. The operating condition matching table uses dust disturbance level as the first search condition and hull sway level as the second search condition; each combination of levels corresponds to a unique metering operating condition number. During processing, the corresponding operating condition table row is first located based on the dust disturbance level, and then the corresponding operating condition table column is located based on the hull sway level. The operating condition number recorded at the intersection of the row and column is the current metering operating condition. The operating condition determination process uses only the combination results of the two levels and does not involve manual selection.

[0161] After determining the current metering condition, the point cloud filtering parameters and dust suppression parameters from the adaptive metering parameter set for the cargo type are read, and the point cloud acquisition parameters and processing parameters are adjusted according to the current metering condition. The point cloud acquisition parameters include the radar scanning frequency and the effective echo reception threshold, while the processing parameters include the time slice length, neighborhood search distance, voxel size, isolated point neighborhood number threshold, and dust point-to-surface distance threshold. Each metering condition corresponds to a set of determined parameter adjustment values, which are stored together with the cargo type's basic parameters in a parameter table.

[0162] In this embodiment, when adjusting the radar scanning frequency, the scanning frequency setting value corresponding to the current metering condition is read and written into the scanning control parameters of the shipborne 3D radar. If the current scanning frequency is inconsistent with the setting value, the current value is replaced by the setting value, and point cloud data is collected according to the new scanning frequency starting from the next complete scanning cycle. The effective echo reception threshold is determined according to the basic echo threshold of the current cargo type and the echo correction amount corresponding to the current operating condition. The sum of the two values ​​is used as the effective echo reception threshold for the current metering cycle. Echoes with a 3D radar reception intensity reaching this threshold are written into the point cloud data, while echoes that do not reach this threshold are not considered valid points for subsequent processing.

[0163] In this embodiment, the time slice length is directly determined based on the time slice setting value corresponding to the current metering condition. When dividing the point cloud in S4, the start and end times of each time slice are continuously determined according to the time slice length, starting from the beginning timestamp of the metering cycle. Therefore, after changes in the ship's sway level, the time slice division can adopt a time scale corresponding to the current condition. The adjustment of the point cloud filtering parameters is based on the basic neighborhood search distance, basic voxel size, and basic isolated point neighborhood number threshold in the current cargo parameter set. The neighborhood distance correction, voxel size correction, and neighborhood number correction corresponding to the current metering condition are read respectively, and each correction is added to the corresponding basic parameters to obtain the actual neighborhood search distance, voxel size, and isolated point neighborhood number threshold used in the current metering cycle. After the above parameters are written, they remain unchanged within the current metering cycle and are used for subsequent point cloud filtering and voxel downsampling of the raster region.

[0164] The dust suppression parameters are adjusted in the same way. First, the basic dust point-to-surface distance threshold and the basic dust neighborhood connection number threshold are read from the current commodity parameter set. Then, the distance correction and connection number correction corresponding to the current metering condition are read. The basic dust point-to-surface distance threshold and the distance correction are added together to obtain the dust point-to-surface distance threshold for the current period; the basic dust neighborhood connection number threshold and the connection number correction are added together to obtain the dust neighborhood connection number threshold for the current period. When S8 performs dust outlier identification, the two adjusted parameters are directly used.

[0165] After all parameters are determined, a parameter record is established using the current metering cycle number. The dust interference level, hull sway level, metering condition number, radar scan frequency, effective echo reception threshold, time slice length, neighborhood search distance, voxel size, isolated point neighborhood number threshold, dust point-to-surface distance threshold, and dust neighborhood connection number threshold are written into this parameter record, and associated with the start and end radar acquisition timestamps of the current metering cycle. Subsequent processing retrieves the parameter record for the corresponding metering cycle based on the radar acquisition timestamp. S4 divides the point cloud according to the time slice length; S5 reads attitude compensation data from the same metering cycle and performs coordinate unification; S6 calculates the grid quality features according to the point cloud processing scale recorded in the current cycle; S7 performs point cloud processing using the corresponding neighborhood search distance, voxel size, and isolated point neighborhood number threshold; S8 uses the dust point-to-surface distance threshold and dust neighborhood connection number threshold to identify dust outliers. After the current measurement cycle ends, the four point cloud quality indicators will be recalculated in the next measurement cycle, and the next set of acquisition parameters and processing parameters will be determined based on the new dust interference level and ship sway level.

[0166] Example 5:

[0167] In this embodiment, S4 specifically includes:

[0168] Extract dry bulk point cloud, radar acquisition timestamp, radar installation pose data, servo stabilization gimbal compensation data, ship attitude data and shore-based reference data from multi-source acquisition data.

[0169] The dry bulk cargo point cloud is sorted by time based on the radar acquisition timestamp, and the time slice length is determined by combining the adjusted point cloud acquisition parameters and processing parameters in S3, thus dividing the dry bulk cargo point cloud into multiple consecutive time slices.

[0170] For each time slice, the corresponding radar installation pose data, servo stabilization gimbal actual compensation attitude, ship attitude data and shore-based reference data are time-aligned.

[0171] Based on the time-aligned data, according to the coordinate transformation relationship between the shipborne 3D radar coordinate system, gimbal coordinate system, ship hull coordinate system and reference coordinate system, the combined coordinate transformation matrix of the corresponding time slice is calculated to correct the coordinate error of the residual pose after servo stabilization.

[0172] Establish a mapping relationship between the combined coordinate transformation matrix and the corresponding time slice, so that the dry bulk cargo point cloud within the same time slice can share the corresponding combined coordinate transformation matrix.

[0173] In the above implementation, S4 reads dry bulk cargo point cloud, radar acquisition timestamp, radar installation attitude data, servo stabilization gimbal compensation data, ship attitude data, and shore-based reference data from the multi-source acquisition data of the current measurement cycle. The dry bulk cargo point cloud is stored according to radar scan frames, with each frame corresponding to one radar acquisition timestamp. The radar installation attitude data records the positional offset of the shipborne 3D radar measurement origin relative to the servo stabilization gimbal installation reference point in three coordinate directions, as well as the three installation rotation angles of the radar coordinate axes relative to the gimbal coordinate axes. The servo stabilization gimbal compensation data uses the actual roll, pitch, and heading compensation angles fed back by the gimbal encoder. The ship attitude data includes roll, pitch, and bow angles. The shore-based reference data records the position coordinates of the ship's reference point relative to a fixed shore-based reference point in three directions. All of the above dynamic data carries the sampling time, with the radar acquisition timestamp serving as the subsequent time correlation reference.

[0174] Specifically, after acquiring point cloud data, the radar acquisition timestamps corresponding to each frame of point cloud within the current measurement period are read, and the point cloud frames are rearranged in chronological order according to the timestamps. When the timestamp values ​​of two point cloud frames are different, the point cloud with the smaller timestamp value is ranked first; when the timestamp values ​​are the same, they are sorted according to the original frame sequence number generated by the radar. This forms a point cloud time sequence consistent with the actual acquisition process, avoiding the influence of data transmission or cache writing order on time slice division. The time slice length is directly read from the adjustment result written to the parameter record of the current measurement period by S3. Specifically, the acquisition time of the first frame of point cloud in the current measurement period is taken as the start time of the first time slice, and one time slice length is added to this start time to obtain the end time of the first time slice; point clouds whose radar acquisition timestamps fall within the range of this start time to end time are included in the first time slice. The second time slice uses the end time of the previous time slice as the starting time, and another time slice length is added to determine the end time, and so on, until all point clouds in the current measurement period have a unique time slice number. When the remaining acquisition time for the last time slice is less than the set time slice length, the end time of the current measurement cycle is taken as the end point of that time slice, and the data from the next measurement cycle is not incorporated into that time slice. A reference time is established for each time slice. The reference time is obtained by dividing the time difference between the start and end times of the time slice by 2, and then adding the resulting time length to the start time of the time slice. The radar installation attitude data is determined by the equipment installation calibration and remains fixed while the mechanical connection between the radar and the gimbal remains unchanged; therefore, it is directly used as the fixed conversion parameter for each time slice. The actual compensated attitude of the servo-stabilized gimbal, the ship's attitude data, and the shore-based reference data are uniformly converted to the reference time of the corresponding time slice.

[0175] Specifically, for the actual attitude compensation of the servo-stabilized gimbal, the data whose sampling time matches the time slice reference time is first retrieved from the gimbal feedback data. If a consistent record exists, the actual roll compensation angle, actual pitch compensation angle, and actual yaw compensation angle in that record are directly read. If no consistent record exists, the feedback data set with the shortest time interval before the reference time and the feedback data set with the shortest time interval after the reference time are selected respectively. The first time interval is obtained by subtracting the first set of feedback times from the reference time, and the second time interval is obtained by subtracting the first set of feedback times from the second set of feedback times. The first time interval is divided by the second time interval to obtain the time position ratio of the reference time between the two gimbal feedbacks. Then, the corresponding angles of the first set are subtracted from the second set of roll, pitch, and yaw angles respectively. The angle difference is multiplied by the time position ratio and then added to the corresponding angles of the first set to obtain the actual roll, pitch, and yaw compensation angles corresponding to that time slice. The ship's attitude data are aligned using the same time reference. When the ship's attitude sensor has a sampling record at the reference time, it directly reads the corresponding roll, pitch, and bow angles. When no corresponding sampling record exists, two sets of ship attitude data before and after the reference time are selected. The time-position ratio is obtained by dividing the time interval between the reference time and the previous set of attitude sampling times by the sampling time interval between the two sets of attitude data. Then, the angle differences of the roll, pitch, and bow angles of the two sets of data are calculated respectively. Each angle difference is multiplied by the time-position ratio and added to the corresponding attitude angle of the previous set to obtain the ship's attitude at the time slice reference time.

[0176] The shore-based reference data determines the position of the ship's reference point according to the time slice reference time. If a shore-based position record exists that matches the reference time, that position record is used directly. If no matching record exists, two sets of position data are obtained, one before and one after the reference time. The difference between the latter set of position coordinates and the former set of position coordinates is calculated for each of the three coordinate directions. This difference is then multiplied by the time-position ratio corresponding to the aforementioned reference time, and the result is added to the former set of position coordinates. The resulting three coordinate values ​​are used as the ship's reference position for that time slice. The position of the shore-based fixed reference point and the direction of the shore-based reference coordinate axes are directly read from the calibration results before measurement and do not change with the time slice.

[0177] Specifically, after time alignment, coordinate transformation relationships are constructed in the order of the shipborne 3D radar coordinate system, gimbal coordinate system, hull coordinate system, and reference coordinate system. The transformation from the radar coordinate system to the gimbal coordinate system is processed first. For any point in the radar point cloud, based on the three installation angles in the radar installation pose data, the three coordinate components of that point are corrected according to the rotation sequence determined during installation calibration. Then, the positional offset of the radar measurement origin relative to the gimbal installation reference point in the three coordinate directions is added. The resulting coordinates are the position of that point in the gimbal coordinate system. Since this transformation relationship is determined solely by the radar installation structure, all time slices under the same installation condition use the same radar installation transformation parameters.

[0178] The transformation from the gimbal coordinate system to the hull coordinate system uses the actual compensated attitude after alignment with the current time slice. First, following the actual mechanical rotation sequence of the gimbal, the spatial orientation of the points is corrected sequentially based on the actual roll compensation angle, pitch compensation angle, and yaw compensation angle. After completing the rotation in the three directions, the fixed position offset of the gimbal rotation center relative to the hull mounting reference point in the three coordinate directions is added, transforming the coordinate relationship of that point from the coordinate relationship that follows the gimbal rotation to the hull coordinate relationship. Here, the actual compensation angle fed back by the encoder is used, not the target compensation angle generated by the servo controller.

[0179] In this embodiment, a transformation from the ship's coordinate system to the reference coordinate system is then performed. Based on the roll, pitch, and bow angles aligned to the current time slice, the spatial orientation of points within the ship's coordinate system is corrected according to the rotation sequence calibrated by the ship's attitude sensors. Then, the ship's reference position, determined by shore-based reference data, is read, and the positional changes of the ship's reference point relative to the fixed shore-based reference point in the three coordinate directions are added to the rotated point coordinates. After this step, the point's position is transformed from a coordinate system that moves with the ship to a reference coordinate system that remains fixed during the measurement process. After the coordinate relationships at each level are determined, the entire parameter calculation process is not repeated sequentially for each radar point; instead, the three-level transformation relationships are pre-combined into a combined coordinate transformation matrix corresponding to the current time slice. During the combination, the radar installation angle and radar installation position offset are first combined with the actual compensation attitude of the gimbal and the fixed installation offset of the gimbal in the order of radar to gimbal and gimbal to hull, to obtain the intermediate transformation relationship from radar coordinate system to hull coordinate system; then the intermediate transformation relationship is further combined with the hull to reference coordinate transformation relationship formed by hull roll, pitch, bow roll and shore reference position.

[0180] In merging position offsets, the position offset generated by the previous level coordinate transformation cannot be directly added to the position offset of the next level. Instead, the direction of the previous level position offset is first transformed according to the actual attitude of the next level coordinate system, and then added to the position offset of the next level itself. When merging rotation relationships, the actual action sequence of radar installation rotation, gimbal actual compensation rotation, and hull attitude rotation is followed to ensure that the combined rotation direction is consistent with the result obtained from the step-by-step coordinate transformation. Through the above processing, a combined coordinate transformation matrix is ​​formed that can transform the radar's original point coordinates to the reference coordinate system in one step.

[0181] The combined coordinate transformation matrix simultaneously undertakes residual pose correction after servo stabilization. The servo-stabilized gimbal performs mechanical compensation based on the target reverse compensation attitude generated by S2, but the actual execution angle is determined by the gimbal encoder feedback. For any time slice, the ship's roll angle and the gimbal's actual roll compensation angle at the same reference moment are transformed to the same rotation direction according to their respective coordinate directions and then superimposed to obtain the residual attitude quantity in the roll direction; the residual attitude quantity in the pitch direction is determined by the ship's pitch angle and the gimbal's actual pitch compensation angle in the same process; the residual attitude quantity in the yaw direction is determined by the ship's yaw angle and the gimbal's actual heading compensation angle. Since the gimbal compensation direction is opposite to the ship's sway direction, when the compensation is completely corresponding, the superposition result after the two are transformed to the same direction is 0. When there is an execution deviation, the superposition result reflects the attitude change in that direction that has not been eliminated by mechanical compensation.

[0182] The remaining attitude parameters are no longer used to establish a second point cloud rotation separately. Instead, they are directly represented by the actual gimbal attitude and the actual ship attitude when generating the ship-to-reference coordinate transformation relationship. The shore-based reference data also provides the actual displacement of the ship relative to the fixed reference position. Therefore, the combined coordinate transformation matrix performs coordinate correction on the angle deviation and ship position change retained after mechanical compensation, so that the point cloud finally falls into the fixed reference coordinate system.

[0183] After calculating the combined coordinate transformation matrix for one time slice, a corresponding data record is created using the time slice number. The record contains the start timestamp, end timestamp, reference time, and combined coordinate transformation matrix for that time slice. When processing subsequent point clouds, the radar acquisition timestamp corresponding to the point cloud is read to determine which time slice's start and end range it falls into, and then the combined coordinate transformation matrix associated with that time slice is called. All point clouds within the same time slice use the same combined coordinate transformation matrix; the gimbal attitude, ship attitude, and shore-based reference data are not reread for individual points.

[0184] Therefore, the continuous point cloud forms a fixed time range according to the time slice length determined by S3. The multi-source pose data within each time slice are converted to the same reference time, and a unique combined coordinate transformation matrix is ​​pre-formed accordingly. When performing coordinate unification in S5, it is only necessary to determine the time slice to which the point cloud belongs based on the radar acquisition timestamp, and then call the corresponding combined coordinate transformation matrix to complete the residual attitude correction and reference coordinate transformation of the point cloud within that time slice.

[0185] Example 6:

[0186] In this embodiment, S5 specifically includes:

[0187] Based on the mapping relationship between time slices and combined coordinate transformation matrices, determine the time slice to which each dry bulk cargo point belongs, and call the corresponding combined coordinate transformation matrix;

[0188] By using a combined coordinate transformation matrix, the dry bulk cargo point cloud within the corresponding time slice is transformed from the shipborne three-dimensional radar coordinate system to the same reference coordinate system, and the initial correction point cloud under each acquisition view is obtained.

[0189] Based on the radar installation pose data and acquisition viewpoint identifiers corresponding to different acquisition viewpoints of one or more shipborne 3D radars and / or the same shipborne 3D radar, the initial correction point cloud is marked with viewpoints and spatial overlapping areas are identified.

[0190] Within the spatially overlapping region, based on the positional deviation, reflection intensity consistency, and point density consistency of the multi-view point cloud, abnormal offset points are removed or their fusion weight is reduced.

[0191] The processed multi-view point clouds are fused, and the time slice identifier, view identifier and fusion weight of each point are retained to generate a corrected point cloud under a unified reference coordinate system.

[0192] In the above implementation, S5 first determines the time slice affiliation of the dry bulk cargo point cloud to be processed based on the mapping relationship between the time slices and the combined coordinate transformation matrix established in S4. The radar acquisition timestamp corresponding to each point is read and compared with the start and end timestamps of each time slice. When the radar acquisition timestamp of a point reaches the start timestamp of a time slice but is less than the end timestamp of that time slice, the point is assigned to that time slice; for the last time slice of the current measurement cycle, the point corresponding to its end timestamp is assigned to that time slice. After the time slice number is determined, the associated combined coordinate transformation matrix is ​​directly read according to the number, ensuring that points within the same time slice use the same set of coordinate transformation parameters. After obtaining the combined coordinate transformation matrix, the dry bulk cargo point cloud within the corresponding time slice is transformed from the shipborne three-dimensional radar coordinate system to the reference coordinate system. For any given point, first read its original coordinate values ​​in the three coordinate directions of the radar coordinate system. Then, perform directional transformation on the three coordinate components according to the pre-merged rotation relationship in the combined coordinate transformation matrix. Finally, superimpose the corresponding position offsets in the three directions to obtain the coordinates of the point in the reference coordinate system. Process all points within the same time slice sequentially, and then process the remaining time slices, thereby forming the initial correction point cloud corresponding to each acquisition viewpoint. This processing directly uses the combined coordinate transformation matrix generated by S4, eliminating the need to repeatedly calculate the radar installation attitude, actual gimbal compensation attitude, ship attitude, and shore-based reference position for individual points.

[0193] For scenarios involving multiple shipborne 3D radars, each radar is pre-assigned a unique device number, and its corresponding radar installation pose data is saved. For cases where the same shipborne 3D radar acquires point clouds from multiple acquisition directions, a corresponding acquisition viewpoint identifier is assigned according to the scanning direction or acquisition position. After the initial calibration point cloud is generated, the radar device number, acquisition viewpoint identifier, and time slice number are written into the data attributes of the corresponding points, ensuring that each point retains its acquisition source after coordinate unification. The spatial overlap region is determined based on the actual coverage of different acquisition viewpoints in the reference coordinate system. For each acquisition viewpoint, the minimum and maximum coordinate values ​​of its initial calibration point cloud in the first and second horizontal directions are calculated to determine the horizontal coverage boundary of that viewpoint. Any two viewpoints are compared; if they share a common coordinate interval in the first horizontal direction and also in the second horizontal direction, the area enclosed by these two common intervals is considered a candidate overlap region.

[0194] Specifically, the candidate overlapping region is further divided according to a set grid size. For each grid, the number of valid points falling into that grid from different viewpoints is counted. When there are valid points from two or more viewpoints within the same grid, that grid is identified as a spatially overlapping grid. Adjacent and contiguous spatially overlapping grids are continuously merged to obtain the spatially overlapping region actually participating in the multi-view consistency judgment. Regions covered by only a single viewpoint do not participate in the multi-view abnormal offset judgment, and their point clouds are directly retained. Within the spatially overlapping region, the positional deviation of the point clouds from different acquisition viewpoints is calculated. Taking the point to be judged in one viewpoint as the target point, the point with the smallest spatial distance in the point set corresponding to another viewpoint is searched as the matching point. When calculating the spatial distance between two points, the coordinate differences of the two points in three coordinate directions are calculated respectively. The three coordinate differences are squared respectively, and the three squared results are added together. The square root of the sum is taken to obtain the positional deviation between the target point and the matching point. When there are three or more acquisition viewpoints, the target point is matched with each of the other viewpoints in the above manner, and the positional deviation of each set is recorded.

[0195] Specifically, the consistency of reflection intensity is determined after spatial matching is completed. First, the reflection intensity of the target point and the matching point is read. For data acquired by different radars, the reflection intensity correction value obtained during the equipment calibration phase is used first. The original reflection intensity is subtracted from the correction value of the corresponding equipment to obtain the reflection intensity under a unified calibration benchmark. Then, the corrected reflection intensity of the target point is subtracted from the corrected reflection intensity of the matching point, and the absolute value of the calculation result is taken as the reflection intensity deviation between corresponding points from two viewpoints. For point clouds acquired by the same radar from different viewpoints, the same difference calculation is performed, but the inter-equipment intensity correction is not applied repeatedly.

[0196] Point density consistency is calculated using spatially overlapping grids. For each overlapping grid, the number of valid points for each acquisition viewpoint within that grid is counted, and then the number of valid points is divided by the horizontal area of ​​the grid to obtain the point density for each viewpoint. When comparing two viewpoints, the difference between the two point densities is first calculated and its absolute value is taken. This absolute difference is then divided by the larger of the two point densities to obtain the point density difference ratio. When there are three or more viewpoints, the point density difference ratio is calculated for each pair of viewpoints. Position deviation, reflection intensity deviation, and point density difference ratio are compared with pre-calibrated position deviation thresholds, reflection intensity deviation thresholds, and point density difference thresholds, respectively. If the position deviation of a target point is less than the position deviation threshold, the reflection intensity deviation is less than the reflection intensity deviation threshold, and the point density difference ratio of its overlapping grid is less than the point density difference threshold, then the target point is determined as a normal overlapping observation point.

[0197] When the positional deviation of a target point reaches a positional deviation threshold, and the reflection intensity deviation also reaches a reflection intensity deviation threshold, the point is listed as a candidate point for abnormal offset. Then, point cloud records of the corresponding spatial locations of this viewpoint in the preceding and following time slices are read. If a candidate point only appears in the current time slice, while points satisfying normal overlapping observation conditions exist in the corresponding locations of both preceding and following time slices, the candidate point is deleted as an abnormal offset point. If a candidate point appears in the same spatial location in consecutive time slices, it is not directly deleted, but instead enters the fusion weight adjustment step to avoid misjudging real surface edges or fixed local protrusions as abnormal points.

[0198] When it is necessary to reduce the fusion weight, a baseline fusion weight is first set for normally overlapping observation points. Then, the deduction amount is determined based on the position deviation, reflection intensity deviation, and point density difference ratio. When the position deviation reaches the corresponding weight correction threshold, a preset position deviation deduction value is subtracted from the baseline fusion weight; when the reflection intensity deviation reaches the corresponding weight correction threshold, a preset reflection intensity deduction value is subtracted; when the point density difference ratio reaches the corresponding weight correction threshold, a preset point density deduction value is subtracted. After completing each deduction, the actual fusion weight of the target point is obtained. When the actual fusion weight is lower than the minimum retention weight specified in the parameter table, it is limited to the minimum retention weight. Each threshold and deduction value is determined during the radar joint calibration process and written into the parameter table.

[0199] In this embodiment, after removing abnormal offset points and adjusting the fusion weights, the multi-view point clouds are fused. For points that do not belong to the spatially overlapping region, their corrected coordinates in the reference coordinate system are directly retained, and their original viewpoint identifier, time slice identifier, and reference fusion weight are recorded. For points that match each other within the spatially overlapping region, the unified spatial position of each point participating in the fusion is calculated according to its respective fusion weight.

[0200] Specifically, for the first coordinate direction, the coordinate values ​​of each matched point in this direction are multiplied by their respective fusion weights, the products are summed, and finally divided by the sum of the weights of all participating points. The result is used as the coordinates of the fused point in the first coordinate direction. The second coordinate direction and the height direction are calculated using the same process, thus obtaining the three coordinate values ​​of the fused point in the reference coordinate system. When using this calculation method, points that have been judged as abnormal and deleted are not included in the weight summation and coordinate calculation. The reflection intensity of the fused point is determined according to the same weight relationship. The reflection intensity of each matched point after equipment correction is multiplied by the corresponding fusion weight, the products are summed, and divided by the sum of all weights participating in the fusion. The result is used as the reflection intensity of the fused point. For points with only one effective viewpoint observation, their coordinates and reflection intensity retain the original values ​​after coordinate correction, and multi-view weighted calculation is not performed. After fusion is completed, a source attribute record is established for each fused point. The source attribute retains all time slice identifiers, all acquisition viewpoint identifiers, and the fusion weights corresponding to each acquisition viewpoint that participated in the formation of the fused point. If a fusion point is formed by three acquisition perspectives, then the three perspective identifiers and three corresponding fusion weights are saved separately, and the source perspectives are not merged into a single identifier; when multiple time slices are involved, the time slice numbers of each source are also retained.

[0201] After the above processing, the point clouds from each acquisition viewpoint first complete the reference coordinate transformation by calling the corresponding combined coordinate transformation matrix according to their respective time slices. Then, in the actual spatial overlapping area, abnormal offset points and fusion weights are determined based on positional deviation, consistency of reflection intensity, and consistency of point density. The processed multi-view data calculates a unified spatial position according to the fusion weights, while retaining the time slice identifier, viewpoint identifier, and corresponding fusion weights. Finally, a corrected point cloud under a unified reference coordinate system is generated and used by S6 for spatial grid division and grid quality level determination.

[0202] Example 7:

[0203] In this embodiment, S6 specifically includes:

[0204] The basic size of the spatial grid cell is determined based on the spatial range of the corrected point cloud, the adaptive metering parameter set for the cargo type, and the adjusted processing parameters.

[0205] The corrected point cloud is divided into multiple spatial grid units according to the basic size, and the time slice identifier, view identifier and fusion weight of the point cloud points in each spatial grid unit are retained.

[0206] Calculate the point density, reflection intensity fluctuation, local height variance, multi-frame repeated observation rate, and multi-view consistency index for each spatial grid cell.

[0207] A raster quality score is generated based on the aforementioned indicators;

[0208] The grid quality level is determined based on the grid quality score, and each spatial grid cell is divided into high-quality grids, medium-quality grids, and low-quality grids.

[0209] In the above implementation, S6 first determines the basic size of the spatial grid corresponding to the correction point cloud. It reads the two horizontal coordinates of all valid correction points in the unified reference coordinate system and obtains the minimum and maximum coordinate values ​​in the two horizontal directions to determine the point cloud coverage of the current measurement area. Points exceeding the pre-calibrated cargo hold boundary are not included in the grid division. Then, it reads the basic grid size corresponding to the current cargo type from the cargo type adaptive measurement parameter set and the adjusted voxel size from the current measurement cycle parameter record formed in S3. It divides the basic grid size by the voxel size, rounds the quotient up, and multiplies the rounded result by the voxel size; the resulting value is used as the basic size of the spatial grid unit. After this processing, the side length of each spatial grid corresponds to an integer voxel side length, avoiding duplicate assignment of voxels after they cross the grid boundary.

[0210] When dividing the calibration point cloud according to the determined base dimensions, the minimum coordinate position of the measurement area in the first and second horizontal directions is used as the starting point for grid division. The grid is then continuously segmented along both horizontal directions according to the base dimensions. For any calibration point, the starting coordinate in the first horizontal direction is subtracted from the coordinate in the first horizontal direction, then divided by the base dimensions, and the integer part of the quotient is taken as the grid number in the first direction. The second horizontal direction is obtained in the same way, and the two numbers together determine the spatial grid cell to which the point belongs. The end grid located at the boundary of the measurement area terminates at the actual boundary, and its area is determined according to the actual coverage area.

[0211] Specifically, when correction points are written to the corresponding spatial raster cells, the time-slice identifier, viewpoint identifier, and fusion weight written in S5 are retained. For points formed by fusing multiple acquisition views, the source information is not merged into a single identifier; instead, the identifiers of each viewpoint involved in the fusion and their corresponding fusion weights are preserved. Data from multiple time slices also retain the corresponding time-slice identifiers. These attributes are then used to calculate the multi-frame repetition rate and multi-view consistency index. For each spatial raster cell, the point density is first calculated. The total number of valid points within the raster is counted, and then the actual area of ​​the raster on the horizontal plane is calculated. The total number of valid points is divided by the actual area, and the result is used as the point density of the raster. For cells located within the complete raster area, the actual area is the product of the base size in two horizontal directions; for cells located at the measurement boundary, the actual horizontal area falling within the valid measurement area is used. This avoids point density distortion caused by incorrect area values ​​for boundary raster cells.

[0212] The reflection intensity fluctuation value is calculated based on the reflection intensity of the valid points within the grid. First, the reflection intensities of all valid points within the grid are summed, then divided by the number of valid points to obtain the average reflection intensity of the grid. Subsequently, the difference between the reflection intensity of each valid point and the average reflection intensity is calculated point by point, and the absolute value of this difference is taken. All absolute values ​​are summed and then divided by the number of valid points; the result is used as the reflection intensity fluctuation value of the current grid. For points from different radar sources, the reflection intensity after intensity correction using S5 is directly used to ensure that the same comparison benchmark is used for different acquisition perspectives.

[0213] In this embodiment, the local height variance is used to characterize the dispersion of the point cloud in the height direction within the grid. During calculation, the height coordinates of all valid points within the grid are first added together, then divided by the number of valid points to obtain the average height of the grid. Then, the difference between the height of each valid point and the average height is calculated point by point. The squared differences are accumulated and finally divided by the number of valid points involved in the calculation to obtain the local height variance of the grid. The obtained value corresponds to the surface roughness characteristic parameters of the current cargo type and is used to distinguish between normal particle undulations and height dispersion exceeding the calibration range of the cargo type. The multi-frame repeated observation rate is calculated based on the time slice identifier. First, the valid points within the current grid are divided into several time slice point sets according to the time slice identifier and arranged in chronological order. For two adjacent time slices, each valid point in the previous time slice is used as a point to be matched, and the point with the smallest spatial distance is searched in the point set of the next time slice. The process of calculating spatial distance is as follows: calculate the coordinate difference between the two points in the three coordinate directions, square the three differences respectively, add the three squared results together, and take the square root of the sum to obtain the spatial distance between the two points.

[0214] When the spatial distance is less than the repeat observation distance threshold specified in the current cargo type parameters, the point to be matched is counted as one repeat observation point. After completing all matching of two adjacent time slots, the repeat observation rate between the two time slots is obtained by dividing the number of repeat observation points by the smaller value of the number of valid points in the two adjacent time slots. When the current grid contains more than three time slots, the above calculation is performed sequentially for each adjacent time slot, and then the repeat observation rates of each group are added together and divided by the number of adjacent time slot groups. The result is used as the multi-frame repeat observation rate of the grid. The multi-view consistency index is determined based on the viewpoint identifier, fusion weight, and spatial correspondence between different viewpoint points. For a grid that contains two or more acquisition views simultaneously, each viewpoint point set is first formed according to the viewpoint identifier, and then point matching is performed on any two views. Taking the point of one viewpoint as the target point, the point with the smallest spatial distance is found in the other viewpoint point set, and the position deviation is obtained by using the same three-direction coordinate difference calculation method as described above.

[0215] After completing the position matching, the corrected reflection intensity of the two matched points is read, the difference between the two reflection intensities is calculated, and the absolute value is taken to obtain the reflection intensity deviation. At the same time, the number of effective points within the grid for each of the two viewpoints is counted, and each is divided by the actual horizontal area of ​​the grid to obtain the point density for the two viewpoints. Then, the absolute value of the difference between the two point densities is calculated, and this absolute value is divided by the larger of the two point densities to obtain the point density difference ratio.

[0216] When the positional deviation of a matching point is less than the positional consistency threshold, the reflection intensity deviation is less than the intensity consistency threshold, and the point density difference ratio of its grid is less than the density consistency threshold, the matching point is counted as a consistent observation point. The consistency rate of the viewpoint combination is obtained by dividing the number of consistent observation points by the number of effective matching points in the viewpoint combination. When there are more than three acquisition views, the consistency rate is calculated for each viewpoint combination, and then the consistency rate of each group is multiplied by the average fusion weight of the corresponding matching point. The resulting multiplications are accumulated and divided by the sum of the average fusion weights of each group. The result is used as the multi-view consistency index of the current spatial grid cell. After obtaining the point density, reflection intensity fluctuation value, local height variance, multi-frame repeated observation rate, and multi-view consistency index, the five indicators are converted into a unified scoring scale. The point density is converted using the benchmark point density corresponding to the current cargo type. The current grid point density is divided by the benchmark point density; when the ratio reaches 1, the point density sub-score is the full score for that item; when the ratio is less than 1, the ratio is multiplied by the full score for that item to obtain the point density sub-score.

[0217] In this embodiment, the reflection intensity fluctuation value is converted using the upper limit of the allowable reflection intensity fluctuation corresponding to the current cargo type. First, the current grid reflection intensity fluctuation value is divided by the upper limit of the allowable fluctuation to obtain the fluctuation ratio. When the fluctuation ratio is less than 1, 1 is subtracted from the fluctuation ratio, and then multiplied by the full score of the item to obtain the reflection intensity sub-score. When the fluctuation ratio reaches 1, the score of this item is recorded as 0. The local height variance is processed according to the upper limit of the allowable height variance corresponding to the current cargo type. The current grid local height variance is divided by the upper limit of the allowable height variance to obtain the height dispersion ratio. When the height dispersion ratio is less than 1, 1 is subtracted from the ratio, and then multiplied by the full score of the item to obtain the local height sub-score. When the height dispersion ratio reaches 1, the score of this item is recorded as 0.

[0218] Both the multi-frame repeat observation rate and the multi-view consistency index are represented by a ratio of 0 to 1. Therefore, the corresponding ratios are directly multiplied by their respective full scores to obtain the multi-frame repeat observation sub-score and the multi-view consistency sub-score. Grids containing only one time slice are not included in the multi-frame repeat observation sub-score, and grids containing only one view are not included in the multi-view consistency sub-score.

[0219] The raster quality score is calculated based on valid sub-scores and their corresponding weights. The parameter table for the current cargo type stores the weights of five indicators: point density, reflection intensity fluctuation, local height variance, multi-frame repeated observations, and multi-view consistency. When all indicators exist, each sub-score is multiplied by its corresponding weight, and the five products are summed to obtain the raster quality score. If any indicators are missing, the weights corresponding to the missing indicators are first deleted, and the original weights of the remaining valid indicators are summed to obtain the total valid weights. Then, the original weights of each valid indicator are divided by this total valid weights to obtain the redistributed weights. The raster quality score is then calculated according to the redistributed weights, ensuring that the sum of the weights actually involved in the calculation remains 1. After the raster quality score is completed, it is compared with two pre-set quality thresholds, where the first quality threshold is greater than the second quality threshold. Spatial raster cells with scores reaching the first quality threshold are classified as high-quality rasters; spatial raster cells with scores below the first quality threshold but reaching the second quality threshold are classified as medium-quality rasters; and spatial raster cells with scores below the second quality threshold are classified as low-quality rasters. The high-quality, medium-quality, and low-quality rasters are grade names derived from determined score ranges and do not rely on manual judgment. After grading, the raster number, five quality indicators, raster quality score, raster quality level, and the time slice identifier, viewpoint identifier, and fusion weight corresponding to each point within the raster are written into the raster attribute record. S7 directly determines the subsequent processing path corresponding to different rasters based on this, without recalculating the aforementioned quality indicators.

[0220] Example 8:

[0221] In this embodiment, S7 specifically includes:

[0222] Low-quality rasters are identified as areas requiring fine-tuning, high-quality rasters as areas requiring simplified-tuning, and medium-quality rasters as areas to be determined.

[0223] A secondary determination is made by combining the quality level of adjacent grids in the area to be determined, the local height variance, and the position of the material pile edge. If the area to be determined is adjacent to a low-quality grid or is located at the edge of the material pile, it is classified into the fine processing area; otherwise, it is classified into the simplified processing area.

[0224] For point cloud points within the fine processing area, calculate local neighborhood density, reflection intensity stability, multi-frame repetition rate, and multi-view consistency, generate point-level credibility scores, and mark real pile surface points, dust outlier points, suspected boundary points, and missing region boundary points;

[0225] For the simplified processing area, the voxel size is determined based on the grid quality score, local height variance and fusion weight of the corresponding spatial grid cell, and voxelization downsampling and lightweight filtering are performed in combination with the point cloud filtering parameters in the cargo type adaptive metering parameter set.

[0226] The point-level reliability assessment results of the fine-processed region and the downsampling results of the simplified region are merged to form a hierarchical point cloud.

[0227] It should be noted that the point-level reliability assessment in this embodiment is a local point cloud processing process after the dust interference level of the current metering cycle has been determined and the parameters have been adjusted in S3. Its purpose is to identify specific dust outliers, real pile surface points and boundary points, and does not re-determine the dust interference level of the current metering cycle.

[0228] In the above implementation, S7 first reads the grid quality level of each spatial grid unit obtained in S6. Spatial grid units with a low quality level are directly assigned to the fine processing area, spatial grid units with a high quality level are directly assigned to the simplified processing area, and spatial grid units with a medium quality level are temporarily stored as areas to be judged. During the division process, the grid number, grid quality score, local height variance, spatial position, and adjacent grid numbers of each grid are retained for subsequent secondary judgment. For areas to be judged, the spatial grid units directly adjacent to it in the four directions (front, back, left, and right) on the horizontal plane are first determined based on the grid number, and the quality level of each adjacent grid is read. If one of the four adjacent grids is a low-quality grid, the current medium-quality grid is directly assigned to the fine processing area. If no low-quality adjacent grids exist, it is then determined whether the current grid is located at the edge of the material pile.

[0229] In this embodiment, the edge of the stockpile is determined by both grid occupancy status and height variation. First, the number of valid stockpile grids and the number of grids without valid stockpile grids in the four adjacent directions of the current grid are counted. If the current grid has valid points and at least one directly adjacent grid does not have valid stockpile points, the current grid is listed as an edge candidate grid. Then, the average height of all valid points in the current grid is calculated, and the average height of adjacent grids with valid stockpile points is calculated for each grid. The average height of the current grid is subtracted from the average height of the adjacent grids, and the absolute value is taken to obtain the height difference in adjacent directions. When the height difference in any direction reaches the edge height determination threshold corresponding to the current cargo type, the grid is determined as a stockpile edge grid and included in the fine processing area.

[0230] Specifically, for medium-quality grids that are neither adjacent to low-quality grids nor identified as the edge of a material pile, the local height variance calculated in S6 is read again and compared with the corresponding height variance threshold in the surface roughness characteristic parameters of the current type of goods. When the local height variance reaches the threshold, the grid is assigned to the fine processing region; when the local height variance is less than the threshold, it is assigned to the simplified processing region. Thus, each spatial grid cell has a unique processing region attribute.

[0231] For each point within the refined processing area, the local neighborhood density is calculated. A 3D neighborhood is established centered on the point to be evaluated, according to the neighborhood search distance specified by the current cargo type's point cloud filtering parameters. For each candidate point within the neighborhood, the coordinate differences between it and the point to be evaluated in three coordinate directions are calculated. The squares of each of the three coordinate differences are then summed, and the square root of the sum is taken to obtain the spatial distance between the two points. Points whose spatial distance is less than the neighborhood search distance are counted as valid neighborhood points. The number of valid neighborhood points is counted and then divided by the spatial volume corresponding to the 3D neighborhood; the result is used as the local neighborhood density of the point to be evaluated.

[0232] Specifically, the stability of reflection intensity is calculated based on the reflection intensity of the point to be evaluated and its effective neighboring points. First, the reflection intensity of all participating points is summed, and then divided by the number of participating points to obtain the local average reflection intensity. Subsequently, the difference between the reflection intensity of each point and the local average reflection intensity is calculated point by point, and the absolute value is taken. Then, all absolute values ​​are summed and divided by the number of points to obtain the local reflection intensity fluctuation value. This fluctuation value is divided by the upper limit of the allowable reflection intensity fluctuation corresponding to the current cargo type to obtain the fluctuation ratio. When the fluctuation ratio is less than 1, the value obtained by subtracting the fluctuation ratio from 1 is used as the base value of reflection intensity stability. When the fluctuation ratio reaches 1, the base value of reflection intensity stability is 0.

[0233] The multi-frame repeat observation rate of the point to be evaluated is determined based on its time-slot identifier. The time-slot number to which the point belongs is read, and the nearest point is searched in the same spatial grid corresponding to the previous and subsequent time-slots. The spatial distance between the matching point and the point to be evaluated is still calculated by summing the squares of the differences in the three coordinate directions, taking the square root. When the spatial distance is less than the repeat observation distance threshold specified for the current cargo type, it is recorded as one valid repeat observation. The multi-frame repeat observation rate of the point is obtained by dividing the number of valid repeat observations by the actual number of adjacent time-slots. When there is only one adjacent time-slot, 1 is used as the comparison count; when there is no adjacent time-slot data, this indicator is marked as missing. Multi-view consistency is determined based on the point's viewpoint identifier and the fusion source retained by S5. When the point to be evaluated has more than two acquisition viewpoint sources, the corresponding observation points are read for any two source viewpoints, the positional deviation between the two observation points is calculated, and the absolute value of the difference in the corrected reflection intensity of the two observation points is calculated. When the positional deviation is less than the positional consistency threshold and the reflection intensity deviation is less than the intensity consistency threshold, the viewpoint combination is recorded as a consistent observation. Multi-view consistency is obtained by dividing the number of consistent observation combinations by the total number of viewpoint combinations actually participating in the comparison. If the point to be evaluated has only one viewpoint source, this indicator is considered missing. After calculating local neighborhood density, reflection intensity stability, multi-frame repeat observation rate, and multi-view consistency, they are converted into point-level sub-scores. When the ratio of local neighborhood density to the baseline neighborhood density of the current cargo type reaches 1, this item receives the set full score; when the ratio is less than 1, the ratio is multiplied by the full score value for this item to obtain the local neighborhood density sub-score. The base value of reflection intensity stability is multiplied by its corresponding full score value to obtain the reflection intensity stability sub-score. The multi-frame repeat observation rate and multi-view consistency are multiplied by their respective corresponding full scores to obtain their respective sub-scores.

[0234] Specifically, the point-level credibility score is calculated based on four sub-scores and their corresponding weights. Each sub-score is multiplied by its weight, and the products are summed to obtain the point-level credibility score for the point being evaluated. If any indicator is missing, its corresponding weight is first deleted, then the original weights of the remaining valid indicators are summed. The original weights of all valid indicators are then divided by the sum of the remaining weights to obtain the redistributed weights, and the point-level credibility score is calculated based on these redistributed weights.

[0235] In this embodiment, the point type is further determined based on the point-level confidence score and spatial location. A point is marked as a true pile surface point when its confidence score reaches the threshold for determining a true pile surface point and it is not located at the edge of the pile or the boundary of a missing point cloud region. A point is marked as a dust outlier point when its confidence score is less than the threshold for determining a dust outlier point, and its local neighborhood density is less than the effective neighborhood density threshold specified for the current cargo type, or its reflection intensity stability is less than the corresponding stability threshold. Points located within the edge grid of the pile undergo a separate boundary preservation judgment. First, the nearest effective pile surface point along the internal direction of the pile is found, and the horizontal distance and height difference between the two points are calculated. Then, the absolute value of the height difference is divided by the horizontal distance to obtain the local slope change value. If the local slope change value does not reach the slope limit corresponding to the current cargo type's stacking angle of repose parameter, and the point forms a valid repeated observation in at least one adjacent time slice, it is marked as a suspected boundary point and is not deleted as a dust outlier point.

[0236] The boundary points of the missing region are determined based on the occupancy of adjacent rasters. When the current raster is directly adjacent to at least one raster without valid stacking points, the completion search distance in the current cargo type point cloud completion parameters is read, and this distance is extended from the common boundary of the rasters into the interior of the current raster to form a boundary band. Points located within the boundary band and whose point-level confidence score reaches the missing boundary retention threshold are marked as missing region boundary points. These points are used as measured boundary constraints for local completion in S8.

[0237] In this embodiment, the simplified processing area does not perform the above-mentioned point-by-point reliability scoring, but first determines the voxel size. The basic voxel size in the current cargo type point cloud filtering parameters is read, along with the current spatial grid cell's grid quality score, local height variance, and fusion weights of each point within the grid. All fusion weights within the grid are summed and then divided by the number of points to obtain the average fusion weight of the grid. A pre-established correspondence between the grid quality score interval, local height variance interval, and average fusion weight interval and the voxel size correction amount is established in the parameter table. The intervals to which the three values ​​of the current grid belong are determined, and the corresponding three voxel correction amounts are read. The basic voxel size is then sequentially added to the three correction amounts to obtain the actual voxel size of the current grid. If the obtained size is less than the minimum voxel size specified for the current cargo type, the minimum voxel size is used; if the obtained size is greater than the maximum voxel size, the maximum voxel size is used. After determining the voxel size, the current simplified processing grid is divided into equal-sized sections along the three coordinate directions. For each point, subtract the starting coordinates of the voxel in the corresponding direction from its three coordinate values, then divide each subtraction by the voxel size. The integer part of the quotient is used as the voxel number in the three directions. Points with the same three voxel numbers are grouped into the same voxel.

[0238] Specifically, for each non-empty voxel, only one representative point is retained. First, the coordinate values ​​of all points within the voxel are added together in three coordinate directions, and then divided by the number of points within the voxel to obtain the average voxel position. Then, the spatial distance between each original point within the voxel and this average position is calculated, and the original point with the smallest spatial distance is selected as the representative point. Using the original point as the representative point, its time slice identifier, view identifier, and fusion weight can still be retained. After voxelization downsampling, lightweight filtering is performed. Centered on each representative point, other representative points are searched according to the filtering neighborhood distance specified by the current cargo type point cloud filtering parameters, and the number of effective neighboring points is counted. If the number of effective neighboring points reaches the minimum retention threshold, the point is retained; if it does not reach the threshold, it is checked whether the point is located in the boundary grid between the fine processing area and the simplified processing area. Points located in the boundary grid and whose spatial distance to the nearest real stack point in the fine processing area is less than the boundary connection distance threshold are retained, and the remaining representative points are deleted.

[0239] After processing, points marked as true pile surface points, suspected boundary points, and missing region boundary points within the finely processed area are merged with representative points retained after voxelization downsampling and lightweight filtering in the simplified processed area. Data already marked as dust outliers in the finely processed area are not included in the merge result. During merging, the original time slice identifier, viewpoint identifier, and fusion weight of each point are retained, and point type and processing area identifiers are added, ultimately forming a hierarchical processed point cloud for S8 to perform dust outlier suppression, boundary preservation, and point cloud missing region completion.

[0240] Example 9:

[0241] In this embodiment, S8 specifically includes:

[0242] Based on the point-level confidence assessment results and the dust suppression parameters in the cargo type adaptive metering parameter set, points with low confidence and low multi-frame repetition rate are marked as dust outliers, points with high confidence are marked as real pile surface points, and points in the middle confidence range are marked as suspected boundary points.

[0243] Dust outliers are removed, while real pile surface points are retained. Suspected boundary points are judged based on the surface roughness characteristic parameters of the corresponding cargo type, and suspected boundary points that conform to the surface geometric undulation characteristics of the corresponding cargo type are retained.

[0244] The missing regions of the point cloud are determined based on the spatial continuity of the actual stacked points and the boundary points of the missing regions, and the missing regions of the point cloud and their adjacent regions are used as local completion regions.

[0245] Within the local completion area, the completion surface is generated based on the stacking angle of repose parameter and point cloud completion parameter in the adaptive metering parameter set for cargo type, combined with local curvature continuity and boundary constraints.

[0246] The completed surface is fused with the preserved real surface points and suspected boundary points to obtain a completed surface point cloud.

[0247] In the above implementation, S8 first reads the point-level confidence score, multi-frame repeat observation rate, point type identifier, and its corresponding spatial grid number output by S7. Simultaneously, it reads the dust outlier determination threshold, the actual stack surface determination threshold, and the repeat observation rate determination threshold from the dust suppression parameters corresponding to the current cargo type. The dust outlier determination threshold is lower than the actual stack surface determination threshold, forming a low confidence interval, an intermediate confidence interval, and a high confidence interval. For each point to be judged, its point-level confidence score is compared with both the dust outlier determination threshold and the actual stack surface determination threshold. When the point-level confidence score is lower than the dust outlier determination threshold, the multi-frame repeat observation rate of that point is read. When the multi-frame repeat observation rate is simultaneously lower than the repeat observation rate determination threshold for the corresponding cargo type, the point is marked as a dust outlier point. When the point-level confidence score reaches the actual stack surface determination threshold, the point is marked as an actual stack surface point. When the point-level confidence score reaches the dust outlier determination threshold and is lower than the actual stack surface determination threshold, the point is marked as a suspected boundary point. Therefore, the determination of dust outliers is constrained by both point-level reliability and time-repeated observation results, avoiding the need to make judgments based solely on the spatial dispersion of a single frame.

[0248] After outlier points are marked, they are removed from the set of points to be reconstructed, while actual pile surface points are directly retained. For suspected boundary points, the surface roughness feature parameters corresponding to the current cargo type are used for further evaluation. Centered on the suspected boundary point, surrounding valid points are extracted according to the neighborhood search distance specified by the surface roughness feature parameters. The average neighborhood height is obtained by adding the height coordinates of all valid points in the neighborhood and dividing by the number of valid points.

[0249] Based on this, the difference between the height of each valid point and the average height of its neighborhood is calculated, and the absolute value of the difference is taken. All absolute values ​​are summed and divided by the number of valid points to obtain the local average height deviation. Simultaneously, the maximum and minimum heights are obtained from the valid neighborhood points, and the minimum height is subtracted from the maximum height to obtain the local height fluctuation. The local average height deviation and local height fluctuation are compared with the corresponding limit values ​​in the surface roughness characteristic parameters of the current cargo type. When neither result meets the corresponding limit value, the suspected boundary point is considered to conform to the normal surface geometric fluctuation characteristics of the current cargo type and is retained; when either result meets the corresponding limit value, the point is not considered a valid boundary point for subsequent completion.

[0250] For suspected boundary points identified through surface roughness, the slope continuity between these points and adjacent actual stack points is checked. The nearest actual stack point is searched, and the horizontal coordinate difference between the two points is calculated. The horizontal distance is obtained by taking the square root of the sum of the squares of the two horizontal coordinate differences. The absolute value of the height coordinate difference between the two points is then calculated, and the local slope is obtained by dividing the height difference by the horizontal distance. This local slope is converted into a slope inclination angle and compared with the current cargo's angle of repose. If the slope inclination angle is less than the angle of repose, the suspected boundary point is retained; if it reaches the angle of repose, the point is removed from the completed boundary constraint points.

[0251] The missing point cloud region is determined jointly based on the spatial continuity of the actual stacked points and the missing region boundary points marked in S7. Following the spatial raster division results of S6, the number of actual stacked points in each spatial raster is counted one by one. If a spatial raster has no actual stacked points, and at least one of its four directly adjacent raster cells (front, back, left, and right) has an actual stacked point, and the raster's edge has a missing region boundary point, then it is listed as a candidate missing raster.

[0252] Specifically, for a missing candidate grid, the nearest real stack surface point is searched along the four directions (front, back, left, and right). If a real stack surface point can be obtained in more than two directions, the horizontal distance and height difference between the real stack surface points in each direction are calculated. The corresponding slope is then obtained by dividing the absolute value of the height difference by the horizontal distance, and the slope is converted into a slope inclination angle. If the slope inclination angle formed by each effective direction is less than the stacking repose angle of the current cargo type, the missing candidate grid is determined as a missing point cloud grid. Adjacent missing point cloud grids sharing a boundary are continuously merged to form a missing point cloud region. Pre-calibrated fixed non-metric areas of the cargo hold are not included in the above missing region determination.

[0253] In this embodiment, after determining the missing point cloud region, based on the adjacency extension distance in the current cargo type point cloud completion parameters, the corresponding distance is extended from the outer boundary of the missing region to the surrounding real stack area. The extended range and the original missing point cloud region are jointly determined as the local completion region. The real stack points, the selected and retained suspected boundary points, and the boundary points of the missing region located within the local completion region are used as completion constraint points. The coordinates of the measured points outside the local completion region remain unchanged. When generating the completion surface within the local completion region, the initial height of the position to be completed is first calculated. For each position to be completed, the nearest valid constraint points on both sides of the position are searched along two horizontal directions, and the height and horizontal distance of each constraint point are recorded. The reciprocal of the horizontal distance from each constraint point to the position to be completed is used as the corresponding distance weight. The height of each constraint point is multiplied by the corresponding distance weight, the products are added together, and finally divided by the sum of all distance weights. The result is used as the initial height of the position to be completed. If there are 4 valid directions, all 4 directions are used in the calculation; if there are 3 valid directions, all 3 directions are used in the calculation; if there are less than 2 valid directions, the complete height for that position is not generated.

[0254] After obtaining the initial height, the slope is checked based on the stacking repose angle parameter. The height difference and horizontal distance between the location to be completed and each effective constraint point are calculated. The local slope is obtained by dividing the absolute value of the height difference by the horizontal distance and converted into a slope inclination angle. When the slope inclination angle in any direction reaches the stacking repose angle for the current cargo type, the allowable height difference is determined by reversing the horizontal distance in that direction and the allowable stacking repose angle. Then, the height of the location to be completed is corrected by adding or subtracting the allowable height difference from the constraint point height, ensuring that the corrected slope inclination angle is less than the stacking repose angle.

[0255] After completing the slope constraints, a local curvature continuity check is performed on the locations to be completed. Three consecutive stack locations are selected along the same horizontal direction. The first local slope (divided by the horizontal distance between the first and second locations) and the second local slope (divided by the horizontal distance between the second and third locations) are calculated. The absolute value of the difference between the first and second local slopes is then calculated to obtain the curvature change value. When the curvature change value reaches the curvature continuity threshold specified by the point cloud completion parameters, the completion height of the middle location is adjusted, and the local slopes on both sides are recalculated until the difference between the local slopes on both sides is less than the curvature continuity threshold. For local completion areas adjacent to cargo hold walls, hold boundaries, or metering boundaries, further boundary constraints are applied. For each completed point generated, its three coordinate values ​​in a unified reference coordinate system are read and compared with the pre-calibrated effective metering boundary of the cargo hold. Completed points located outside the effective metering boundary are deleted; those located within the effective metering boundary are retained. Simultaneously check whether the height direction is within the allowable measurement space of the cargo hold; completion points exceeding the allowable height range are not included in the completion surface. After all the positions to be completed satisfy the stacking angle of repose constraint, curvature continuity condition, and boundary constraint, the completion points are connected according to spatial adjacency to form a local completion surface. Positions lacking two effective constraint directions, or those that still cannot satisfy the curvature continuity condition after height correction, are not generated as completion points to avoid extrapolation at positions lacking measured constraints.

[0256] Finally, the completed surface is merged with the retained real surface points and suspected boundary points. For missing locations where no measured points exist, the corresponding completed points of the completed surface are directly used. For locations where the completed surface and measured points fall into the same spatial grid, real surface points and suspected boundary points are retained first, and measured points are not replaced by completed points. When setting a transition zone at the edge of the missing region, the transition weight is determined based on the horizontal distance from the completed point to the boundary of the missing region. The completion weight is obtained by dividing this distance by the width of the transition zone, and then the measured weight is obtained by subtracting the completion weight from 1. The measured height is multiplied by the measured weight, and the completed height is multiplied by the completion weight, and the two products are added together. The result is used as the fusion height of the transition position. Beyond the transition zone, the height of the completed surface after constraint is directly used inside the missing region. After fusion, the real surface points, suspected boundary points that conform to the surface geometric undulation characteristics of the current cargo type, and effective completed points are uniformly combined to form a completed surface point cloud, and the measured source identifier and the completed source identifier are retained respectively. The resulting completed point cloud is used as input data for S9 to perform adaptive surface reconstruction and volume calculation.

[0257] Example 10:

[0258] In this embodiment, S9 specifically includes:

[0259] The local height variance, local curvature, and point cloud reliability of each region of the completed pile surface are used to determine the local undulation level of the pile surface, combined with the location of the pile edge.

[0260] The surface reconstruction grid size range is determined based on the grid reconstruction parameters in the adaptive metering parameter set for cargo type, and the surface reconstruction grid size corresponding to each region is determined within the grid size range based on the local undulation level of the stack surface.

[0261] Based on the determined surface reconstruction mesh size, the point cloud of the completed stack surface is adaptively meshed to generate a surface model of the dry bulk cargo stack surface;

[0262] The volume calculation reference plane is determined based on the ship's hold bottom model and shore-based reference data, and the height difference between the dry bulk cargo stack surface model and the volume calculation reference plane is calculated.

[0263] The volume is integrated based on the area and corresponding height difference of each grid cell to obtain the dry bulk cargo volume measurement results, and the grid size, point cloud confidence level and the proportion of the completed area for each region are output.

[0264] In the above implementation, S9 uses the completed pile surface point cloud obtained in S8 as the processing object and divides the local processing area according to the spatial grid position relationship established in S6. Each local processing area retains the height coordinates, point-level confidence level, measured source identifier, completed source identifier, and pile edge identifier of each point, which are used for subsequent undulation level determination and surface reconstruction parameter determination. The local height variance is calculated according to the height distribution of the effective points in the current area. First, the height coordinates of all effective points in the area are added together, and then divided by the number of effective points to obtain the average height of the area. Then, the difference between the height of each effective point and the average height is calculated sequentially, the squares of each difference are accumulated, and the accumulated result is divided by the number of effective points. The resulting value is used as the local height variance of the area. This value directly reflects the dispersion of each point in the current area around the average pile surface height. The local curvature is determined according to the slope change of adjacent pile surface positions. Three consecutive grid positions are selected sequentially along the first horizontal direction. The height of the first position is subtracted from the height of the second position, and then divided by the horizontal distance between the two positions to obtain the first slope. The height of the second position is then subtracted from the height of the third position, and then divided by the corresponding horizontal distance to obtain the second slope. The absolute value of the subtraction of the first slope from the second slope is taken as the curvature change value of this set of positions in the first horizontal direction. The same process is repeated along the second horizontal direction to calculate the corresponding curvature change value. The curvature change values ​​in the two directions are then added together and divided by 2 to obtain the local curvature of that position. The local curvature of all computable positions within the current region is summed and divided by the number of positions to obtain the local curvature of the entire region.

[0265] Specifically, point cloud reliability is calculated based on the reliability of each point within the region and the data source. For actual stack points and retained suspected boundary points, the point-level reliability scores generated by S7 and S8 are directly read. For completed points generated by S8, the completion reliability coefficient is read from the point cloud completion parameters of the current cargo type. The point-level reliability scores of each measured point within the region and the completion reliability coefficients of each completed point are added together, and then divided by the total number of points involved in the calculation. The result is used as the point cloud reliability of the region. Thus, the regional point cloud reliability reflects both the reliability of the measured points and the occupancy of the completed data within the region.

[0266] The level of local undulation on the pile surface is determined based on local height variance, local curvature, point cloud confidence, and the position of the pile edge. The mesh reconstruction parameters of the current cargo type pre-save two height variance boundary values, two local curvature boundary values, and one point cloud confidence boundary value. The second height variance boundary value is greater than the first height variance boundary value, and the second local curvature boundary value is greater than the first local curvature boundary value.

[0267] When the local height variance of the current region is less than the first height variance threshold and the local curvature is less than the first local curvature threshold, it is initially determined to be at level 1 undulation. When the local height variance reaches the second height variance threshold, or the local curvature reaches the second local curvature threshold, it is initially determined to be at level 3 undulation. All other cases are determined to be at level 2 undulation. After the initial determination, the undulation level is corrected based on the location of the stockpile edge and the point cloud reliability. If the current region contains a suspected boundary point retained by S8, or if the region is directly adjacent to a region without effective stockpile points, it is determined to be a stockpile edge region. If the initial level of the stockpile edge region is level 1 undulation, it is corrected to level 2 undulation; if the initial level is level 2 undulation, it is corrected to level 3 undulation. When the point cloud reliability of the current region is less than the point cloud reliability threshold, the existing undulation level is also increased by one level; if it is already at level 3 undulation, it remains unchanged. After the above processing, each local region obtains a unique local undulation level for the stockpile surface.

[0268] Specifically, the range of grid sizes for surface reconstruction is determined based on the grid reconstruction parameters corresponding to the current cargo type. The grid reconstruction parameters store the minimum grid size, the reference grid size, and the maximum grid size, with the minimum grid size being smaller than the reference grid size, and the reference grid size being smaller than the maximum grid size. The minimum grid size is used for the third undulation level, the reference grid size for the second undulation level, and the maximum grid size for the first undulation level. Therefore, areas where the changes in stack height and curvature meet the criteria for the third undulation level are reconstructed using the minimum size, while areas for the first undulation level are reconstructed using the maximum size. When two adjacent areas use different grid sizes, a transition grid is established at the common boundary of the two areas using the grid size with the smallest value, ensuring that the grid nodes on both sides of the boundary have corresponding connection positions. The transition process only changes the grid division scale and does not move the measured coordinates of the actual stack points and suspected boundary points.

[0269] After determining the grid size for each region, adaptive meshing is applied to the completed point cloud. The minimum coordinate values ​​of the current region in the two horizontal directions are used as the starting point for partitioning, and the region is continuously partitioned along the two horizontal directions according to the corresponding grid size. For each point, the region's starting coordinates in the corresponding direction are subtracted from its two horizontal coordinates, and then divided by the current grid size. The integer part of the quotient is used as the grid number in the two directions, and the two grid numbers determine the surface reconstruction mesh element to which the point belongs.

[0270] For each non-empty mesh cell, the reconstructed height is determined based on the height and confidence level of its valid points. First, the height of each point within the mesh is multiplied by its corresponding confidence level to obtain a weighted height value. Then, all weighted height values ​​are summed, along with all confidence levels used in the calculation. The sum of these weighted height values ​​is divided by the sum of the confidence levels, and the result is used as the reconstructed height of that mesh cell. Real surface points, suspected boundary points, and completion points all participate in the above calculation using their respective confidence levels. For meshes without valid points, the reconstructed height is not directly generated. When the mesh is within the range of the valid completion surface already formed by S8, the height of the corresponding completion surface is read as the reconstructed height of the mesh; when it is not within the range of the valid completion surface, the mesh remains empty, and extrapolation based on distant mesh heights is not performed.

[0271] After all valid meshes have been reconstructed to their heights, the mesh nodes are connected according to their adjacency relationships. For an area enclosed by four adjacent mesh nodes, it is divided into two triangular patches along a pre-fixed diagonal direction. For each triangular patch, the height difference and horizontal distance between the nodes at both ends of its edge are checked, and the corresponding slope is determined by dividing the absolute value of the height difference by the horizontal distance. The slope inclination angle is then determined based on the slope. When the slope inclination angle caused by the completion point reaches the current cargo's angle of repose, the completion constraint result of S8 is used to re-limit the height of the completion point. Slope changes formed by actual stacking points maintain the original measured height and are not flattened. After processing all valid meshes in sequence, a dry bulk cargo stacking surface model is formed.

[0272] Specifically, the reference surface for volume calculation is determined based on the ship's hold hull model and shore-based reference data. The ship's hold hull model is established before metrology operations, recording the 3D coordinates of each model node in the ship's hull coordinate system. During volume calculation, the position and attitude of the ship relative to the reference coordinate system are read from the shore-based reference data, and the nodes of the hull model are transformed from the ship's hull coordinate system to a unified reference coordinate system according to the coordinate transformation sequence used in S5. After transformation, the hull model and the dry bulk cargo surface model are on the same coordinate reference. For each surface reconstructed mesh, the corresponding reference height is determined in the hull model according to its horizontal position. When the center position of the mesh coincides with the horizontal position of a node in the hull model, the height of that node is directly used as the reference height. When the grid center is located within the area enclosed by four adjacent hull model nodes, the horizontal distances from the grid center to the four hull nodes are calculated respectively, and the reciprocal of each horizontal distance is used as the distance weight. The heights of the four hull nodes are multiplied by their corresponding distance weights, summed, and then divided by the sum of the four distance weights. The resulting height is used as the reference height for volume calculation at that grid location.

[0273] The height difference of each grid cell is then calculated. The effective stacking height of the current grid is obtained by subtracting the bilge reference height at the same horizontal position from the reconstructed stacking height. If the effective stacking height is greater than 0, the grid is included in the volume integration; if the effective stacking height is equal to 0, the local volume of the grid is recorded as 0; if the result is less than 0, the grid is checked against the ship's structural boundary, and grids belonging to non-cargo spaces are removed from the volume calculation range. The local volume of each grid cell is calculated based on the actual area of ​​the grid and the effective stacking height. The area of ​​a complete grid is obtained by multiplying the side lengths of two horizontal grids, and then multiplying this area by the corresponding effective stacking height; the result is used as the local volume of that grid. For incomplete grids located at the cargo hold boundary, the horizontal area is determined based on the actual overlap between the grid boundary and the effective measurement boundary, and then multiplied by the corresponding effective stacking height. The local volumes of all effective grids are summed sequentially, and the result is used as the dry bulk cargo volume measurement result.

[0274] When saving the surface model of the stack surface in the form of triangular patches, the difference between the stack surface height at each of the three vertices of the triangular patch and the corresponding hull reference height is calculated. The three height differences are added together and divided by 3 to obtain the average effective height of the triangular patch. Then, the horizontal projected area of ​​the triangular patch is determined based on the horizontal coordinates of the three vertices. Multiplying the horizontal projected area by the average effective height yields the local volume of the triangular patch. The local volumes of all triangular patches are added together to obtain the volume result for the same measurement region. After volume integration, the surface reconstruction mesh size, point cloud confidence level, and completed region proportion used for each local region are output simultaneously. The surface reconstruction mesh size is directly read from the final mesh size used in that region; the point cloud confidence level uses the point cloud confidence level calculation result for the aforementioned region; the completed region proportion is determined based on the completed region source identifier.

[0275] When calculating the proportion of the completed area, first count the number of valid reconstructed grids containing completed points within the current area, then count the total number of valid grids participating in surface reconstruction within the area. Divide the number of valid grids containing completed points by the total number of valid reconstructed grids, and the result is taken as the proportion of the completed area for that region. For the case where the same grid contains both measured points and completed points, it is counted as a grid containing completed points, as long as a completed point exists within that grid. Finally, the total volume of dry bulk cargo, the surface reconstruction grid size corresponding to each local area, the point cloud reliability, and the proportion of the completed area are written into the current metrology task results. Thus, the final volume is calculated from the actual height difference between the surface of the cargo bed and the hull datum under a unified reference coordinate system, while retaining the reconstruction scale and data source information for each local area.

[0276] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive and accurate measurement of multiple cargo types in dry bulk cargo based on three-dimensional radar point clouds, characterized in that, Includes the following steps: S1. Acquire multi-source data from the dry bulk cargo metering area and determine the adaptive metering parameters corresponding to the current cargo type based on the cargo type information. S2. Determine the ship's sway level based on multi-source acquired data, and control the servo stabilization gimbal to perform reverse attitude compensation for the shipborne three-dimensional radar. S3. Determine the dust interference level based on the point cloud quality index of multi-source data, and adaptively adjust the point cloud acquisition parameters and processing parameters in conjunction with the ship sway level and cargo type adaptive metering parameters. S4. Divide the dry bulk cargo point cloud into multiple time slices according to the multi-source acquisition data, and pre-calculate the combined coordinate transformation matrix for residual attitude correction for each time slice. S5. Based on the combined coordinate transformation matrix, the dry bulk cargo point clouds from different acquisition perspectives are uniformly transformed to the same reference coordinate system to obtain the correction point cloud; S6. Divide the corrected point cloud into multiple spatial grid cells and determine the grid quality level based on the point cloud quality characteristics of each spatial grid cell. S7. Determine the fine processing region and the simplified processing region based on the raster quality level. Perform point-level confidence assessment on the fine processing region and perform adaptive voxelization downsampling and lightweight filtering on the simplified processing region. S8. Based on the point-level credibility assessment results and the adaptive metering parameters of the cargo type, suppress dust outliers, retain real pile surface points and suspected boundary points, and only perform physical constraint completion on the missing area of ​​the point cloud and its adjacent area to obtain the completed pile surface point cloud. S9. Based on the completed point cloud of the stack surface, the adaptive metering parameters of the cargo type, and the degree of local stack surface undulation, adaptive surface reconstruction and volume calculation are performed to obtain the volume measurement results of dry bulk cargo.

2. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 1, characterized in that, The multi-source data collected in S1 includes dry bulk cargo point cloud, radar acquisition timestamp, radar installation attitude data, servo stabilization gimbal compensation data, ship attitude data, shore-based reference data, and cargo type information. Based on the cargo type information, a cargo type adaptive measurement parameter set is generated. The cargo type adaptive measurement parameter set includes point cloud filtering parameters, dust suppression parameters, point cloud completion parameters, stacking angle of repose parameters, surface roughness feature parameters, and mesh reconstruction parameters. The cargo type adaptive measurement parameter set is used to adapt parameters for point cloud processing, dust discrimination, boundary preservation, local completion, and surface reconstruction processes corresponding to different cargo types.

3. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 2, characterized in that, S2 specifically includes: Extract hull attitude data and radar acquisition timestamps from multi-source data, and synchronize the hull attitude data according to the radar acquisition timestamps. The amplitude of the hull attitude change is calculated based on the synchronized hull attitude data, and the hull rolling level is determined based on the amplitude of the hull attitude change. The target reverse compensation attitude is determined based on the synchronized hull attitude data, and the compensation gain, compensation rate and compensation limit parameters of the servo stabilization gimbal are determined based on the hull sway level. Based on the target's reverse compensation attitude and the compensation gain, compensation rate, and compensation limiting parameters, a gimbal compensation command is generated, causing the servo stabilization gimbal to drive the shipborne three-dimensional radar to perform attitude compensation opposite to the direction of the ship's sway. Obtain the actual compensated attitude of the servo-stabilized gimbal and write the hull sway level and actual compensated attitude into the timing compensation data.

4. The adaptive and accurate measurement method for multiple types of dry bulk cargo based on three-dimensional radar point clouds according to claim 2, characterized in that, S3 specifically includes: Dry bulk cargo point clouds are extracted from multi-source data, and the point density, reflection intensity fluctuation, point cloud hole ratio, and multi-frame repeated observation rate of the dry bulk cargo point clouds are calculated. The dust interference level is determined based on the point cloud quality indicators. Match the dust disturbance level with the hull sway level to determine the current metering conditions; Based on the current metering conditions and the point cloud filtering parameters and dust suppression parameters in the adaptive metering parameter set for the cargo type, adjust the point cloud acquisition parameters and processing parameters. Write the adjusted point cloud acquisition and processing parameters into the current metering cycle so that subsequent time slice division, coordinate unification, raster quality grading, and point cloud processing are performed according to the adjusted parameters.

5. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 2, characterized in that, S4 specifically includes: Extract dry bulk point cloud, radar acquisition timestamp, radar installation pose data, servo stabilization gimbal compensation data, ship attitude data and shore-based reference data from multi-source acquisition data. The dry bulk cargo point cloud is sorted by time based on the radar acquisition timestamp, and the time slice length is determined by combining the adjusted point cloud acquisition parameters and processing parameters, thus dividing the dry bulk cargo point cloud into multiple consecutive time slices. For each time slice, the corresponding radar installation pose data, servo stabilization gimbal actual compensation attitude, ship attitude data and shore-based reference data are time-aligned. Based on the time-aligned data, according to the coordinate transformation relationship between the shipborne 3D radar coordinate system, gimbal coordinate system, ship hull coordinate system and reference coordinate system, the combined coordinate transformation matrix of the corresponding time slice is calculated to correct the coordinate error of the residual pose after servo stabilization. Establish a mapping relationship between the combined coordinate transformation matrix and the corresponding time slice, so that the dry bulk cargo point cloud within the same time slice can share the corresponding combined coordinate transformation matrix.

6. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 5, characterized in that, S5 specifically includes: Based on the mapping relationship between time slices and combined coordinate transformation matrices, determine the time slice to which each dry bulk cargo point belongs, and call the corresponding combined coordinate transformation matrix; By using a combined coordinate transformation matrix, the dry bulk cargo point cloud within the corresponding time slice is transformed from the shipborne three-dimensional radar coordinate system to the same reference coordinate system, and the initial correction point cloud under each acquisition view is obtained. Based on the radar installation pose data and acquisition viewpoint identifiers corresponding to different acquisition viewpoints of one or more shipborne 3D radars and / or the same shipborne 3D radar, the initial correction point cloud is marked with viewpoints and spatial overlapping areas are identified. Within the spatially overlapping region, based on the positional deviation, reflection intensity consistency, and point density consistency of the multi-view point cloud, abnormal offset points are removed or their fusion weight is reduced. The processed multi-view point clouds are fused, and the time slice identifier, view identifier and fusion weight of each point are retained to generate a corrected point cloud under a unified reference coordinate system.

7. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 6, characterized in that, S6 specifically includes: The basic size of the spatial grid cell is determined based on the spatial range of the corrected point cloud, the adaptive metering parameter set for the cargo type, and the adjusted processing parameters. The corrected point cloud is divided into multiple spatial grid units according to the basic size, and the time slice identifier, view identifier and fusion weight of the point cloud points in each spatial grid unit are retained. Calculate the point density, reflection intensity fluctuation, local height variance, multi-frame repeated observation rate, and multi-view consistency index for each spatial grid cell. A raster quality score is generated based on the aforementioned indicators; The grid quality level is determined based on the grid quality score, and each spatial grid cell is divided into high-quality grids, medium-quality grids, and low-quality grids.

8. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 7, characterized in that, Specifically, S7 includes: Low-quality rasters are identified as areas requiring fine-tuning, high-quality rasters as areas requiring simplified-tuning, and medium-quality rasters as areas to be determined. A secondary determination is made by combining the quality level of adjacent grids in the area to be determined, the local height variance, and the position of the material pile edge. If the area to be determined is adjacent to a low-quality grid or is located at the edge of the material pile, it is classified into the fine processing area; otherwise, it is classified into the simplified processing area. For point cloud points within the fine processing area, calculate local neighborhood density, reflection intensity stability, multi-frame repetition rate, and multi-view consistency, generate point-level credibility scores, and mark real pile surface points, dust outlier points, suspected boundary points, and missing region boundary points; For the simplified processing area, the voxel size is determined based on the grid quality score, local height variance and fusion weight of the corresponding spatial grid cell, and voxelization downsampling and lightweight filtering are performed in combination with the point cloud filtering parameters in the cargo type adaptive metering parameter set. The point-level reliability assessment results of the fine-processed region and the downsampling results of the simplified region are merged to form a hierarchical point cloud.

9. The adaptive and accurate measurement method for multiple types of dry bulk cargo based on three-dimensional radar point clouds according to claim 8, characterized in that, S8 specifically includes: Based on the point-level confidence assessment results and the dust suppression parameters in the cargo type adaptive metering parameter set, points with low confidence and low multi-frame repetition rate are marked as dust outliers, points with high confidence are marked as real pile surface points, and points in the middle confidence range are marked as suspected boundary points. Dust outliers are removed, while real pile surface points are retained. Suspected boundary points are judged based on the surface roughness characteristic parameters of the corresponding cargo type, and suspected boundary points that conform to the surface geometric undulation characteristics of the corresponding cargo type are retained. The missing regions of the point cloud are determined based on the spatial continuity of the actual stacked points and the boundary points of the missing regions, and the missing regions of the point cloud and their adjacent regions are used as local completion regions. Within the local completion area, the completion surface is generated based on the stacking angle of repose parameter and point cloud completion parameter in the adaptive metering parameter set for cargo type, combined with local curvature continuity and boundary constraints. The completed surface is fused with the preserved real surface points and suspected boundary points to obtain a completed surface point cloud.

10. The adaptive and accurate measurement method for multiple cargo types of dry bulk cargo based on three-dimensional radar point clouds according to claim 9, characterized in that, S9 specifically includes: The local height variance, local curvature, and point cloud reliability of each region of the completed pile surface are used to determine the local undulation level of the pile surface, combined with the location of the pile edge. The surface reconstruction grid size range is determined based on the grid reconstruction parameters in the adaptive metering parameter set for cargo type, and the surface reconstruction grid size corresponding to each region is determined within the grid size range based on the local undulation level of the stack surface. Based on the determined surface reconstruction mesh size, the point cloud of the completed stack surface is adaptively meshed to generate a surface model of the dry bulk cargo stack surface; The volume calculation reference plane is determined based on the ship's hold bottom model and shore-based reference data, and the height difference between the dry bulk cargo stack surface model and the volume calculation reference plane is calculated. The volume is integrated based on the area and corresponding height difference of each grid cell to obtain the dry bulk cargo volume measurement results, and the grid size, point cloud confidence level and the proportion of the completed area for each region are output.

Images