A Smart Detection Method and System for Ship Dimensions and Load Condition Based on Multi-Frame Heading Correction

Abstract

This invention discloses an intelligent detection method and system for ship dimensions and loading status based on multi-frame heading correction. This method, without requiring camera intrinsic and extrinsic parameter calibration or image instance segmentation masks, estimates the global vertical direction using radar data and constructs a static background raster channel mask. After limiting the search domain, it performs density clustering and fragment merging on the water point cloud. A dual-algorithm optimal orientation strategy is used on the horizontal projection plane to determine the heading envelope to measure the length, width, and height of a single frame. Within the ship's horizontal footprint, the mode of the height histogram is used to estimate the freeboard, and the draft and displacement are calculated by combining the registered depth or a global regression model based on length and width, thereby inferring the loading status. Finally, the navigation frame sequence is rotated using double-angle vector averaging and fused with frame center alignment to refine the output dimensions. This invention overcomes technical challenges such as unilateral obstruction, water surface interference, and longitudinal registration degradation on smooth hulls, achieving stable and accurate measurement of ship dimensions and loading status at fixed monitoring stations throughout the day.

Description

Technical Field

[0001] This invention belongs to the field of water traffic perception and ship monitoring technology, and in particular, it is a method and system for intelligent detection of ship size and load condition based on multi-frame heading correction. Background Technology

[0002] Vessel monitoring in inland waterways and near-shore channels requires obtaining vessel dimensions, draft, and loading status information through non-contact methods. Fixed-point deployed lidar can continuously collect point clouds of passing vessels, offering advantages such as all-weather operation and remote deployment. In existing technical solutions, one approach relies on multi-sensor fusion or camera-radar joint calibration, using image segmentation masks to filter projected point clouds to improve hull separation accuracy. However, this method has high calibration and maintenance costs and cannot adequately meet the monitoring deployment requirements that necessitate lidar alone to output loading status indicators.

[0003] Another type of method directly performs geometric clustering and bounding box measurement on single-frame point clouds, which is simple to implement, but has three prominent defects in fixed site application scenarios: First, static structures such as riverbanks, utility poles, and distant tree lines are present in almost every frame of point cloud; second, duckweed, algae, and water surface echoes are spatially close to the ship hull; and third, lidar usually only illuminates one side of the ship, and the waterline is often mistakenly deleted because it is at the same height as the water surface.

[0004] Furthermore, draft and load information are located below the waterline and need to be estimated using registered depth or statistical regression methods. While existing technologies suggest establishing depth regression models based on ship type clusters, during online measurement, radar typically only reliably acquires ship length and beam data. The accuracy gains from clustering routing must be weighed against system complexity. Additionally, separately calculating background occupancy for a single slow-moving vessel point cloud sequence can easily misclassify hull echoes as static structures, leading to their removal in subsequent processing. Fixed checkpoints on inland waterways also need to monitor traffic flow, suspected overloading, and other regulatory indicators.

[0005] In existing technologies, frame-by-frame estimation of the vertical direction leads to tilting of the horizontal projection; voxel-based background subtraction methods are prone to erosion of the hull in waterways with frequent ship traffic; multi-frame longitudinal cross-correlation registration techniques suffer from aperture ambiguity on smooth hulls; and the shape-depth regression method based on ship type offers limited accuracy gains when only ship length and beam data are available during online measurement. Therefore, constructing a pure radar geometry processing pipeline that requires no camera calibration, operates stably at fixed waterway checkpoints, and outputs load condition monitoring indicators is a pressing technical challenge. Summary of the Invention

[0006] The purpose of this invention is to address the deficiencies or shortcomings of the existing technology by providing a method and system for intelligent detection of ship size and loading status based on multi-frame heading correction.

[0007] The technical solution to achieve the purpose of this invention is: an intelligent detection method for ship size and loading status based on multi-frame heading correction, the method comprising the following steps:

[0008] Step 1: Obtain the lidar point cloud data collected by the fixed monitoring station, and estimate the global vertical direction vector based on the multi-frame point cloud to establish a horizontal projection coordinate system;

[0009] Step 2: In the horizontal projection coordinate system, a background occupancy model is constructed by statistically analyzing the frequency of occurrence of spatial grids in point clouds of multiple frames or multiple ship trips, and a channel grid mask is constructed using non-static grids to limit the ship hull search area.

[0010] Step 3: Estimate the water surface height based on the point cloud within the channel grid mask, filter candidate points with a preset margin above the water surface, perform clustering and fragment merging, and obtain the hull point set;

[0011] Step 4: Project the hull point set onto the horizontal plane, and use a dual-algorithm optimal orientation strategy to determine the ship's heading envelope, thereby obtaining the ship's length, width, and height in a single frame;

[0012] Step 5: Within the determined horizontal footprint range of the ship, extract sampling points from the complete point cloud, estimate the main deck height based on the statistical characteristics of the height histogram of the sampling points relative to the water surface height distribution, and calculate the freeboard;

[0013] Step 6: Obtain the ship's molded depth, calculate the draft based on the molded depth and freeboard, and then estimate the displacement and determine the ship's loading status.

[0014] Step 7: For the navigation frame sequence of the same ship, perform derotation and center alignment fusion using the heading angle and envelope geometric center of each frame, and refine the output ship size on the fused point cloud.

[0015] Furthermore, in step 1, the global vertical direction vector is obtained by fitting it onto the dominant near-horizontal plane of the scene using a random sampling consensus algorithm, and a constant normal constraint is maintained for all point cloud frames collected by the monitoring station.

[0016] Furthermore, in step 2, the background occupancy model is specifically a static background occupancy grid model in a horizontal projected coordinate system:

[0017] The occurrence frequency is defined by the number of frames, that is, the proportion of the number of frames in which the horizontal grid cell appears in the single frame point cloud of multiple shipments is used as the occupancy rate, and the same grid cell is counted only once in the same frame, so that the occupancy rate reflects the probability of the grid cell appearing in the frame sequence rather than the point cloud density, thereby achieving decoupling between the occupancy rate and the point cloud density.

[0018] Grids with an occupancy rate not lower than a preset threshold are marked as static backgrounds. The channel grid mask is constructed using non-static grids, and no deletion-style background subtraction is performed on the point cloud within the ship hull search area defined by the channel grid mask.

[0019] Furthermore, in step 3, the clustering segmentation adopts a density-based spatial clustering algorithm; the fragment merging specifically involves: using qualified clusters whose point counts meet the hard constraints of ship size as hull seed clusters, iteratively merging spatially adjacent fragment clusters with horizontal distances within a preset span to restore the truncated hull line; the hard constraints of ship size include at least: ship length range, ship width range, upper limit of cluster height, and lower limit of length-to-width ratio.

[0020] Furthermore, when estimating the water level height in step 3, a slow tidal correction is also included: the trend function of water level slowly changing over time is fitted using the water level quantile sequence of the same monitoring station within a preset historical time window, and the water level height is dynamically corrected according to the current timestamp during online measurement to eliminate the influence of intraday river level fluctuations on the water level candidate point screening threshold and freeboard calculation.

[0021] Furthermore, the dual-algorithm optimal orientation strategy described in step 4 is specifically as follows:

[0022] Simultaneously, the minimum bounding rectangle of the hull point set on the horizontal plane and the principal axis-aligned bounding rectangle based on principal component analysis are obtained;

[0023] Using the ship's slender geometry prior as the criterion for heading correctness, when the ratio of the aspect ratio of the principal axis-aligned circumscribed rectangle to the aspect ratio of the smallest circumscribed rectangle is greater than a preset selection coefficient, the principal axis-aligned circumscribed rectangle is selected as the final ship heading envelope; otherwise, the smallest circumscribed rectangle is selected. The preset selection coefficient is greater than 1.

[0024] Furthermore, the calculation of freeboard in step 5 is specifically as follows:

[0025] The sampling points are taken from the points in the complete point cloud within the horizontal footprint of the ship, and the extraction range of the sampling points is decoupled from the water surface candidate point screening threshold used for clustering and segmentation in step 3.

[0026] After removing points within the preset waterline height, a relative height histogram is calculated. The relative height corresponding to the bin center, which represents the mode peak in the height histogram, is used as the estimated value of the main deck height, and thus as the estimated value of the freeboard.

[0027] Furthermore, in step 6, the method for obtaining the ship's shape depth includes: firstly querying the registration database by ship identification code; if the query fails, then using the measured ship length and width, predicting the shape depth through a single global shape depth regression model;

[0028] The single global deep regression model includes:

[0029] When only the ship's length and width are known online, a second-order polynomial model using only the ship's length and width as input is adopted;

[0030] Given the deadweight tonnage and gross tonnage, a log-linear power-law model is used, which includes the ship's length, width, deadweight tonnage, and gross tonnage as inputs.

[0031] Furthermore, the rotation removal and center alignment fusion described in step 7 does not rely on nearest neighbor registration between point clouds, and its specific implementation method is as follows:

[0032] For the heading angle of each frame in the navigation frame sequence Taking the major axis without direction, the reference heading for a single voyage is obtained by averaging the angled vectors:

[0033]

[0034] Frame by frame, the frame center determined by the envelope geometry is used as the translation anchor point for center alignment. Rotation is then performed by combining the heading angle of each frame to avoid aperture degradation issues in longitudinal registration of smooth hulls. Time add Eliminate ambiguity regarding 180° orientation;

[0035] Before derotation and center alignment fusion, abnormally wide frames are removed using the low quantile of the single-frame ship width distribution in each frame.

[0036] On the fused point cloud, the length and width of the ship are refined and output along the preset quantile boundary range of the ship's length axis and width axis.

[0037] Furthermore, the specific criterion for removing abnormally wide frames is: calculating the ship width of each frame. lower quantiles Only retain those that meet the requirements. The frames are used to avoid the problem of incorrectly removing normal wide-body cargo ships due to the contaminated frames raising the mean or median caused by the right skew of the distribution; where m represents the unit meter.

[0038] Furthermore, before outputting the dimensions and load conditions in step 7, multi-level verification and anomaly correction steps are included:

[0039] The measured ship dimensions are then subjected to size range tests, length-to-beam ratio tests, and regression matching tests of ship beam to ship length, as well as registered Ku Bayesian posterior corrections.

[0040] The Bayesian posterior correction of the registration library is specifically as follows: taking the conditional distribution of the registration library as the prior, when the measured ship width deviates from the conditional expectation by more than a preset multiple of the standard deviation of the registration library conditions, the Bayesian posterior expectation of the prior and the measured ship width is calculated, and the measured ship width is replaced by the posterior expectation as the final output.

[0041] Furthermore, after determining the load condition in step 6, a consistency check of the displacement is also included:

[0042] The geometric displacement of the ship is calculated and compared with the displacement estimated based on the power law of the registry. When the relative deviation between the two exceeds a preset deviation threshold, the confidence level of the load condition determination is reduced or the ship is marked as suspected of being overloaded.

[0043] Furthermore, the method also includes a single navigation time window delineation step:

[0044] The ship is visually tracked based on camera images synchronized with the lidar. The start and end times of a single navigation passage from the ship's entry into the field of view to its departure are determined to define the navigation time window. All lidar point cloud frames with corresponding timestamps within the navigation time window are captured to form the navigation frame sequence. The visual tracking does not rely on camera intrinsic and extrinsic parameter calibration and image instance segmentation mask.

[0045] A space-velocity joint feature model is constructed for the navigation frame sequence. The radial velocity of the ship is fitted using the horizontal footprint center and timestamp of each frame. The fusion weight of each frame is determined by combining the radial velocity with the point cloud coverage of each frame. Frames that are close to the horizontal plane and have high coverage are assigned greater weights to suppress the dilution of the fusion envelope by sparse frames at the ends during multi-frame fusion.

[0046] Furthermore, the method also includes a quality scoring step: synthesizing a comprehensive quality score identifier based on the multi-frame hull point cloud coverage, multi-frame freeboard consistency, dimensional anomaly detection results, and displacement consistency verification results; when the quality score identifier is lower than a preset quality threshold, a low confidence mark is output or a manual review process is triggered.

[0047] On the other hand, a ship size and loading condition intelligent detection system based on multi-frame heading correction is provided for executing the aforementioned ship size and loading condition intelligent detection method based on multi-frame heading correction. The system includes:

[0048] The first module is used to acquire lidar point cloud data collected by fixed monitoring stations and organize single or multi-frame sequences according to navigation events.

[0049] The second module is used to estimate the global vertical direction vector, establish the horizontal projection coordinate system, and construct the static background occupancy model and the channel grid mask.

[0050] The third module is used to estimate the water level height and perform geometric clustering segmentation and neighbor fragment merging on the point cloud within the mask-defined search domain.

[0051] The fourth module is used to measure the size of a single frame using a dual-algorithm optimal orientation strategy, and to estimate the freeboard within the horizontal footprint of the hull using a height histogram.

[0052] The fifth module is used to calculate the draft, estimate the displacement, and determine the loading status by combining the global regression model or the registered database.

[0053] The sixth module is used to perform multi-frame heading derotation and center alignment fusion on the same navigation sequence, and to perform multi-level verification correction or measurement quality scoring.

[0054] Compared with the prior art, the significant advantages of this invention are:

[0055] (1) The present invention adopts pure radar geometric measurement and load condition detection as the main path. Without the need for synchronous camera internal and external parameter calibration or the need to call image instance segmentation mask, it can achieve stable measurement of ship size and load condition at fixed checkpoints, which greatly reduces the system deployment complexity (preferably a single global regression model) and maintenance cost.

[0056] (2) This invention constructs a global static background grid and channel mask offline using single-frame data from multiple ships, effectively limiting the ship search area and eliminating interference from riverbanks and fixed structures. Simultaneously, it defines occupancy rate as a statistical indicator based on frame count rather than point count, and stipulates that slow-moving sequences of individual ships cannot be statistically analyzed separately, effectively avoiding the failure mode where large, slow-moving ships are mistakenly identified as static backgrounds and eliminated. Furthermore, it replaces voxel-point deletion-based background subtraction with channel grid masks, replaces one-dimensional cross-correlation longitudinal registration with frame center motion compensation, and outputs draft, displacement, load conditions, and checkpoint navigation statistics, allowing deployment without visual calibration.

[0057] (3) Based on the physical characteristics of the dense horizontal echo layer formed on the main deck of a ship, this invention performs height histogram mode estimation on the complete point cloud data within the horizontal projection range of the hull to obtain freeboard data. This method is completely decoupled from the clustering algae filtering threshold, effectively avoiding the systematic clamping error generated by the traditional low quantile method for ships with low freeboard.

[0058] (4) In the process of horizontal envelope estimation, this invention uses the minimum area bounding rectangle (MAR) and principal component analysis (PCA) principal frame together, and uses the "length-to-width ratio geometric prior" of the ship's slender shape as the basis for determining the correctness of orientation. This design effectively alleviates the problem of the ship's width being too large when the point cloud is sparse or has an L-shaped distribution, which is caused by the frame falling into the diagonal direction due to orientation misjudgment.

[0059] (5) In the multi-frame fusion stage, this invention uses the “frame center” determined by the hull envelope geometry of each frame as the translation anchor point, replacing the traditional nearest neighbor matching method; at the same time, it uses the double-angle vector averaging method to deal with the problem of ambiguity in the major axis orientation. This design effectively avoids the multi-frame size divergence caused by longitudinal registration degradation (aperture problem) in the nearest neighbor registration process (such as the ICP algorithm) of smooth hulls, and significantly improves the integrity of hull size information.

[0060] (6) This invention constructs a multi-level verification chain that includes size range verification, aspect ratio verification, regression matching verification, and Bayesian posterior correction of the registry library. By calculating the posterior expectation through a log-normal prior and performing smooth correction, it can not only robustly eliminate or correct abnormal frames caused by interference from adjacent ships and shore walls, but also effectively reduce the risk of normal wide-body cargo ships being incorrectly corrected.

[0061] (7) This invention can not only stably output ship length, ship width and ship height, but also further calculate draft, displacement and load status (empty / half-loaded / full-loaded) through global regression model and multi-level verification. It can also introduce tidal slow change correction and navigation flow volume aggregation statistics, providing a strong regulatory-level digital basis for waterway volume analysis and overload suspicion screening.

[0062] (8) This invention can select a single navigation time window based on synchronous camera visual tracking, capture all point cloud frames from the ship's entry to departure, and construct a space-velocity joint feature model to characterize the relationship between radial velocity and heading changes and point cloud coverage differences; accordingly, fusion weights are assigned to each frame to suppress the dilution effect of sparse end frames on the fusion envelope in scenarios with drastic changes in radial velocity. This technical approach does not rely on camera intrinsic and extrinsic parameter calibration and instance segmentation masks, and is compatible with pure radar geometry master paths.

[0063] (9) The present invention further integrates point cloud coverage, freeboard consistency, size anomaly detection and displacement consistency verification into a quality identifier, so that regulatory business can distinguish between automatic reliable results and low confidence results that require manual review.

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0065] Figure 1This is a schematic diagram of a method for intelligent detection of ship size and loading condition based on multi-frame heading correction in one embodiment.

[0066] Figure 2 This is a functional block diagram of a ship size and loading condition intelligent detection system based on multi-frame heading correction in one embodiment.

[0067] Figure 3 This is a schematic diagram of the multi-frame heading correction and center alignment fusion process in one embodiment.

[0068] Figure 4 This is a schematic diagram of a preferred deep regression model in one embodiment.

[0069] Figure 5 This is a schematic diagram comparing an embodiment with existing technical solutions.

[0070] Figure 6 This is a schematic diagram of the visual tracking navigation time window delineation and the space-velocity joint feature model in one embodiment. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0072] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0073] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0074] In one embodiment, combined Figure 1 This paper provides an intelligent detection method for ship size and loading status based on multi-frame heading correction. The method includes the following steps:

[0075] Step 1: Obtain the lidar point cloud data collected by the fixed monitoring station, and estimate the global vertical direction vector based on the multi-frame point cloud to establish a horizontal projection coordinate system;

[0076] Step 2: In the horizontal projection coordinate system, a background occupancy model is constructed by statistically analyzing the frequency of occurrence of spatial grids in point clouds of multiple frames or multiple ship trips, and a channel grid mask is constructed using non-static grids to limit the ship hull search area.

[0077] Step 3: Estimate the water surface height based on the point cloud within the channel grid mask, filter candidate points with a preset margin above the water surface, perform clustering and fragment merging, and obtain the hull point set;

[0078] Step 4: Project the hull point set onto the horizontal plane, and use a dual-algorithm optimal orientation strategy to determine the ship's heading envelope, thereby obtaining the ship's length, width, and height in a single frame;

[0079] Step 5: Within the determined horizontal footprint range of the ship, extract sampling points from the complete point cloud, estimate the main deck height based on the statistical characteristics of the height histogram of the sampling points relative to the water surface height distribution, and calculate the freeboard;

[0080] Step 6: Obtain the ship's molded depth, calculate the draft based on the molded depth and freeboard, and then estimate the displacement and determine the ship's loading status.

[0081] Step 7: For the navigation frame sequence of the same ship, perform derotation and center alignment fusion using the heading angle and envelope geometric center of each frame, and refine the output ship size on the fused point cloud.

[0082] The fusion scheme based on camera hull mask and point cloud projection relies on pre-calibrated intrinsic and extrinsic parameters, focusing on cross-frame whole-ship point cloud stitching and high-precision geometric locking; while the present invention is aimed at the deployment scenario without camera calibration, with waterway grid mask, geometric clustering and freeboard-loading condition inference chain as the core, the two are independent of each other in terms of necessary input, hull locking path and regulatory output indicators.

[0083] Furthermore, in one embodiment, in step 1, the global vertical direction vector is obtained by fitting it onto the dominant near-horizontal plane of the scene using a random sampling consensus algorithm, and a constant normal constraint is maintained for all point cloud frames collected by the monitoring station.

[0084] Specifically:

[0085] Point sets are obtained by parsing the 3D coordinates of point cloud files acquired by LiDAR. The ship's deck is approximately horizontal, and the normal to the dominant near-horizontal plane can be used as an estimate of the gravity direction. The Random Sample Consensus (RANSAC) algorithm is used to fit the plane equation on the dominant near-horizontal plane. Retain the normal constraint (normal component) that is consistent with the sensor mounting direction. satisfy The candidate normal of ) at the interior point threshold Maximize the number of interior points to obtain the global unit normal. ,in The unit normal vector of the fitted plane (which, after normalization, satisfies...) ), This represents the three-dimensional coordinate vector of any point in the point cloud participating in the fitting. This represents the constant term in the plane equation.

[0086] Construct a horizontal orthogonal basis: ,in To and Non-collinear preset reference vectors. The horizontal projected coordinates of a point are represented as... Vertical height is expressed as This global vertical direction vector maintains a constant constraint across all frames at all stations.

[0087] Preferably, in some embodiments, All frames at the monitoring stations remain constant.

[0088] Here, if the vertical direction is estimated separately for each frame, the horizontal projection will be tilted, and the measurements of the ship's length and width may be significantly distorted (for example, from approximately 48m × 12m to approximately 140m × 76m). Therefore, the present invention prefers a globally unified estimation. .

[0089] It should be noted that in specific scenarios where an ideal near-horizontal plane is lacking, global vertical orientation estimation can also be achieved by extracting vertical normal alignment from known vertical buildings in the scene (such as bridge piers and utility poles); or, in the case of a radar rigidly equipped with a three-axis accelerometer / IMU, spatial attitude transformation alignment can be directly achieved by using the gravity vector output by the hardware.

[0090] Furthermore, in one embodiment, in step 2, the background occupancy model is specifically a static background occupancy grid model in a horizontal projected coordinate system:

[0091] The occurrence frequency is defined by the number of frames, that is, the proportion of the number of frames in which the horizontal grid cell appears in the single frame point cloud of multiple shipments is used as the occupancy rate, and the same grid cell is counted only once in the same frame, so that the occupancy rate reflects the probability of the grid cell appearing in the frame sequence rather than the point cloud density, thereby achieving decoupling between the occupancy rate and the point cloud density.

[0092] Grids with an occupancy rate not lower than a preset threshold are marked as static backgrounds. The channel grid mask is constructed using non-static grids, and no deletion-style background subtraction is performed on the point cloud within the ship hull search area defined by the channel grid mask.

[0093] Specifically:

[0094] Set the horizontal grid side length (e.g.) In the horizontal projection coordinate system, a static background model is constructed by statistically analyzing the horizontal grid occupancy rate of point clouds from multiple ship-related single-frame analyses. To avoid artificially inflated grid occupancy rates caused by high-density ship echoes at close range, the occupancy rate is defined by frame count rather than point count: for each grid... Its occupancy rate is defined as ,in For the first The set of grid cells in which a frame appears, with the same grid cell counted only once within the same frame. The proportion of frames appearing in this set must not be lower than a preset threshold (e.g., ...). The raster markers are designated as static backgrounds. A non-static raster is used to construct a channel raster mask to constrain the clustering search domain, and no deleting background subtraction is performed on the point cloud within the channel mask. The background model must be constructed offline from single-frame data from multiple ship trips; individual statistics for slow-speed sequences of a single ship are not permitted.

[0095] It should be noted that the occupancy rate definition reflects how many frames a grid appears in, rather than being dominated by point cloud density. This avoids the possibility that the occupancy rate of the grid containing the hull will be artificially high due to the high density of the hull echo at close range, and thus be mistakenly judged as a static background. This is different from the background subtraction method of occupying grids by accumulating the number of points.

[0096] It should be noted that the two-dimensional square grid can be replaced with a hexagonal honeycomb grid, a polar coordinate system fan grid, or a non-uniformly divided octree spatial voxel grid, depending on the waterway topography, as long as the criterion of "defining the occupancy rate by frame number and offline statistics for multiple ships" is met.

[0097] Furthermore, in one embodiment, in step 3, the clustering segmentation adopts a density-based spatial clustering algorithm; the fragment merging specifically involves: using qualified clusters whose point counts meet the ship size hard constraints as hull seed clusters, iteratively merging spatially adjacent fragment clusters with horizontal distances within a preset span to restore the truncated hull line; the ship size hard constraints include at least: ship length range, ship width range, upper limit of cluster height, and lower limit of length-to-width ratio.

[0098] Specifically:

[0099] Estimate the current water level based on the point cloud within the channel mask. , This represents the height value of point p. For binary mask functions, Indicates whether point p is located within the waterway area. This means filtering out all results that meet the criteria. The height value of point p The set that constitutes.

[0100] Select the preset margin above the water surface (Preferred) And below the maximum height limit (Preferred) candidate point set .

[0101] Perform density-based spatial clustering algorithms (such as DBSCAN, with the optimal parameter being the neighborhood radius) on the candidate point set. Minimum number of samples (While meeting hard constraints on ship dimensions, including at least: length range, beam range, upper limit of cluster height, and lower limit of length-to-beam ratio, such as length...) , ship width Cluster height From the qualified clusters, the cluster with the most points and the densest density is selected as the hull seed cluster. Subsequently, the horizontal spatial distance is iteratively adjusted within a preset span threshold (e.g., ...). The spatially adjacent fragment clusters within the seed cluster are merged into the seed cluster, forming a complete hull point set.

[0102] Furthermore, in one embodiment, step 3, when estimating the water level height, also includes a tidal slow-varying correction: a trend function of water level slowly changing over time is fitted using the water level quantile sequence of the same monitoring station within a preset historical time window, and the water level height is dynamically corrected based on the current timestamp during online measurement to eliminate the influence of intraday river level fluctuations on the water surface candidate point screening threshold and freeboard calculation; the specific formula is expressed as:

[0103]

[0104] In the formula, The water level quantile sequence within a preset historical time window at the same checkpoint is fitted to correct the impact of intraday river level fluctuations on the algae filtration threshold. The real-time corrected water level at time t. It is the 12th percentile. For the set of waypoints.

[0105] Here, although the inland river checkpoint is not a seaport, the water level can still fluctuate slightly by 0.1–0.4 meters daily (due to flood discharge and reservoir capacity management). In fixed... Based on this, fitting , Take 12.4h or 24h. Estimated from the low quantile amplitude of the channel points over the past 7 days.

[0106] The algae filtering threshold, i.e., the candidate point screening threshold on the water surface, is changed to Freeboard calculation is used synchronously This correction reduces the probability of low-freeboard vessels being mistakenly judged as half-loaded when water levels rise in the afternoon.

[0107] Furthermore, in one embodiment, the dual-algorithm optimization orientation strategy in step 4 specifically refers to:

[0108] Simultaneously, the minimum bounding rectangle of the hull point set on the horizontal plane and the principal axis-aligned bounding rectangle based on principal component analysis are obtained;

[0109] Using the ship's slender geometry prior as the criterion for heading correctness, when the ratio of the aspect ratio of the principal axis-aligned circumscribed rectangle to the aspect ratio of the smallest circumscribed rectangle is greater than a preset selection coefficient, the principal axis-aligned circumscribed rectangle is selected as the final ship heading envelope; otherwise, the smallest circumscribed rectangle is selected. The preset selection coefficient is greater than 1.

[0110] Specifically:

[0111] Project the locked set of hull points onto the horizontal plane, and simultaneously execute two envelope algorithms: one is to use the minimum bounding rectangle (MAR) to obtain the aspect ratio. Secondly, the covariance matrix is ​​calculated based on principal component analysis (PCA), and the principal axis-aligned bounding rectangle is extracted to obtain the aspect ratio. The system uses the ship's slender "length-to-width ratio geometric prior" as the criterion for correctness: when the relation is satisfied... ( A preset selection coefficient greater than 1 (preferably 1.15) is used when the principal axis-aligned bounding rectangle is selected; otherwise, the bounding rectangle with the smallest area is selected. This selection strategy can effectively suppress the error of excessive ship width caused by the smallest area rectangle falling in the diagonal direction under sparse point clouds. The horizontal footprint range of the ship's hull in a single frame is determined by the envelope of the selected rectangle. .

[0112] It should be noted here that the selection coefficient is... The selection is based on the geometric prior of the ship's length-to-beam ratio: the ship's length-to-beam ratio usually belongs to When the smallest circumscribed rectangle falls in the diagonal direction, its aspect ratio is... Approaching 1 contradicts the aforementioned geometric prior; therefore, the aspect ratio is used as the criterion for orientation correctness, rather than the envelope fitting residual, so that the optimal decision is consistent with the ship's slender geometric prior, which is different from existing methods that use the fitting residual or the minimum area of ​​multiple models as the optimal criterion.

[0113] It should be noted that the dual-algorithm optimal combination is not limited to MAR and PCA. It can be equivalently replaced by extracting the longest side line based on the stochastic Hough Transform as the main axis orientation, and using the historical frame Kalman filter state as a priori direction gating constraint for optimal decision-making.

[0114] Furthermore, in one embodiment, calculating the freeboard in step 5 specifically involves:

[0115] The sampling points are taken from the points in the complete point cloud within the horizontal footprint of the ship, and the extraction range of the sampling points is decoupled from the water surface candidate point screening threshold used for clustering and segmentation in step 3.

[0116] After removing points within the preset waterline height, a relative height histogram is calculated. The relative height corresponding to the bin center, which represents the mode peak in the height histogram, is used as the estimated value of the main deck height, and thus as the estimated value of the freeboard.

[0117] Specifically:

[0118] To avoid the systematic constraint of algae filter margin on low freeboard vessels, freeboard estimation is decoupled from the clustering of algae filter subsets in step 3. The system extracts the complete point cloud within the horizontal footprint of the hull. Points within the area are used as height sampling points. Extremely low heights near the waterline (such as...) are excluded. After the points within the range, calculate the height histogram of the relative water surface height (e.g., relative height). Take the lower part The histogram has a bin width of 0.4m. Based on the physical prior that a dense horizontal echo layer forms on the main deck of a ship, the maximum peak (mode peak) in the height histogram corresponds to the main deck height. Therefore, the relative height value corresponding to the center of the mode bin in the height histogram is extracted as the estimated value of the main deck height. The freeboard was calculated. .

[0119] It should be noted that the freeboard estimation uses the mode of the height histogram instead of the low quantile. The principle is as follows: the main deck at the bottom of the hull forms a dense horizontal echo layer, which is represented by the mode peak in the height histogram. Although the waterline is at a lower height, it has been filtered out by the water surface debris removal threshold. Although the cargo stack is at a higher height, the number of points is dispersed. Therefore, the mode corresponds to the height of the main deck, while the low quantile corresponds to the filtered waterline remnants or noise, which do not constitute an effective estimate of the freeboard. This estimation is decoupled from the clustering algae filtering threshold to avoid the systematic constraint of algae filtering margin on ships with low freeboard.

[0120] Furthermore, in one embodiment, step 6, the method of obtaining the ship's shape depth includes: firstly querying the registration database by ship identification code; if the query fails, then using the measured ship length and width, predicting the shape depth through a single global shape depth regression model;

[0121] The single global deep regression model includes:

[0122] When only the ship's length and width are known online, a second-order polynomial model using only the ship's length and width as input is adopted;

[0123] Given the deadweight tonnage and gross tonnage, a log-linear power-law model is used, which includes the ship's length, width, deadweight tonnage, and gross tonnage as inputs.

[0124] Specifically:

[0125] Deep Priority is given to querying the registration database by vessel identification code. If the query fails and the deadweight tonnage (DWT) and gross tonnage (GT) are known, a log-linear power-law model is used for prediction; if only the vessel's length L and width W are known online, a second-order polynomial global regression model is used to predict the depth. .

[0126] Preferably, the log-linear power-law model is as follows: , For the intercept term; Let L be the regression coefficient for the ship's length. The regression coefficients are for deadweight tons (DWT). The regression coefficient is the width W of the ship. is the regression coefficient for gross tonnage (GT).

[0127] Preferably, the second-order polynomial global regression model is: ,in, Leave room for collection .

[0128] Calculate the draft based on the molded depth and freeboard. Then estimate the geometric displacement. (where the square coefficient is) water density Ultimately, the ratio of draft to depth is used to determine the final determination. Determine the ship's condition: To be fully loaded, The first is unloaded, the rest are half-loaded.

[0129] It should be noted that when the registration database lacks explicit ship type labels, K-Means clustering based on scale features can be performed, followed by cluster-based deep regression fitting. Clustering after standardization The contour coefficient is preferred. On the 80% / 20% holdout set: when using all feature routes, the clustered second-order polynomial RMSE is 0.597m, and the global second-order polynomial RMSE is 0.638m;

[0130] However, when only the ship's length L and width W are available online, The clustered RMSE of the route is 0.628m, which is only about 1.6% better than the global model, and small sample clusters require regression to the global model. Therefore, this invention preferably uses a single global L, W second-order polynomial as the default deep estimator; when MMSI or DWT, GT are available, database lookup or enabling a full-feature log-linear model is still preferred. See also Figure 4 .

[0131] Furthermore, in one embodiment, the derotation and center alignment fusion described in step 7 does not rely on nearest neighbor registration (such as ICP / NDT) between point clouds, in order to avoid aperture ambiguity caused by longitudinal registration degradation of smooth hulls. Combined with Figure 3 The specific implementation method is as follows:

[0132] Step 7-1: Set the heading angle of each frame in the navigation frame sequence. Taking the major axis without direction, the reference heading for a single voyage is obtained by averaging the angled vectors:

[0133]

[0134] Step 7-2, frame by frame, determine the frame center using the envelope geometry. Center alignment is performed using the translation anchor points, and rotation is then performed by combining the heading angles of each frame. To avoid the aperture degradation problem in longitudinal registration of smooth hulls, and when Time add Eliminate ambiguity regarding 180° orientation; among which, Let R be the new coordinates of the point after rotation and translation, and let R be the two-dimensional rotation matrix. The original coordinates;

[0135] Step 7-3: Before rotation removal and center alignment fusion, use the low quantile of the single-frame ship width distribution of each frame to remove abnormally wide frames.

[0136] Step 7-4: On the fused point cloud, along the preset quantile boundary range of the ship's length axis and width axis. The output should specify the length and width of the vessel, where... Preset cropping positions, For lower quantile ( (Quantity) boundary values.

[0137] Preferably, in some embodiments, the freeboard is estimated for each frame. Remove For abnormal frames, perform a heading de-rotation and frame center alignment to reduce the contamination of the merged envelope by individual misselected clusters; among which The preset threshold for identifying abnormal frames is set.

[0138] Preferably, in some embodiments, the specific criterion for removing abnormally wide frames is: calculating the ship width of each frame. lower quantiles Only retain those that meet the requirements. The frames are selected to avoid the problem of incorrectly removing normal wide-body cargo ships due to the right skew of the distribution causing the mean or median to be raised by contaminated frames.

[0139] It should be noted here that the abnormally wide frame removal is done at the low quantile. As a robust estimate of the true ship beam, the underlying principle is that pollution from adjacent vessels or shorelines always makes the ship beam appear larger rather than smaller, thus the beam distribution is skewed to the right and at the lower quantile. The true width of the uncontaminated ship is reflected in the frame width, while the mean or median will be inflated by contaminated frames; therefore, Retaining frames avoids the accidental removal of normal wide-body cargo ship frames, unlike methods that remove abnormal frames using a fixed threshold or average.

[0140] It should be noted that the frame center alignment and fusion is based on the principle of addressing the aperture problem of a smooth ship hull: displacement along the longitudinal direction of the smooth ship hull. At this time, the nearest neighbor distance of each point remains approximately unchanged, that is, the nearest neighbor registration cost is... (Constant), therefore, the registration of nearest neighbor class point clouds degenerates in the vertical direction;

[0141] This invention replaces nearest neighbor matching with the orientation of the frame center as the translation anchor point. The frame center is determined by the hull envelope geometry of each frame, which does not rely on point cloud matching, thereby avoiding longitudinal registration failure caused by aperture issues. This is different from multi-frame point cloud registration methods that aim to minimize the nearest neighbor distance.

[0142] Furthermore, in one embodiment, before outputting the dimensions and load conditions in step 7, a multi-level verification and anomaly correction step is included:

[0143] The measured ship dimensions are then subjected to size range tests, length-to-beam ratio tests, and regression matching tests of ship beam to ship length, as well as registered Ku Bayesian posterior corrections.

[0144] The Bayesian posterior correction of the registry specifically involves learning prior relations from the registry: ,in To predict the ship's beam; based on the distribution of registration conditions. As a priori, when the actual measured ship width Deviation from conditional expectation Exceeding the preset multiple Standard deviation of registration database conditions At that time, calculate the Bayesian posterior expectation of the prior and the measured ship width, replace the measured ship width with the posterior expectation as the final output, and record the replacement fact in the verification log.

[0145] Furthermore, the multi-level verification chain is specified, and the following steps are performed sequentially on the measured ship length L, ship width W, and height H:

[0146] First-level interval test, Second-level aspect ratio test; The third-level regression matching test of ship beam to ship length is performed by fitting the registry database. (Approximately 120,000 samples) ),Require ;

[0147] The fourth-level registration library uses Bayesian posterior corrections based on a log-normal conditional distribution. a priori ( (where the standard deviation is 1), calculate the posterior expectation of the ship's beam for the first three trigger levels. ,when (Preferred) ) time replace .

[0148] The correction triggers a Level 5 displacement consistency review: based on the corrected... Geometric draft calculation , with the power law of the registered reservoir for drainage Comparison, when relative deviation (Preferably, the deviation threshold) If the measurement quality score is lowered, the air traffic will be marked as suspected of being overloaded or requiring manual review.

[0149] Each level of triggering and replacement is recorded in the verification log, retaining the original measurement values, correction values ​​at each level, and trigger reasons to meet regulatory requirements for traceability and auditability. Compared with single-level hard truncation, multi-level chains do not trigger subsequent corrections at normal wide-body cargo ships and converge level by level at abnormal wide frames, balancing the strength of corrections with the risk of erroneous corrections.

[0150] Furthermore, in one embodiment, after determining the load condition in step 6, a displacement consistency check is also included:

[0151] Calculate the geometric displacement of a ship And the displacement estimated based on the power law of the registry. A comparison is made, and when the relative deviation between the two exceeds a preset deviation threshold (i.e., ... Preferably When the value is 0.35, the confidence level of the load condition determination is reduced or the vessel is marked as suspected of being overloaded.

[0152] Furthermore, in one embodiment, the method further includes a single navigation time window delineation step:

[0153] The vessel is visually tracked based on camera images synchronized with the lidar to determine the start and end times of a single passage from the vessel's entry into the field of view to its departure. Define a navigation time window and capture all lidar point cloud frames with corresponding timestamps within the navigation time window to form the navigation frame sequence, wherein the visual tracking does not rely on camera intrinsic and extrinsic parameter calibration and image instance segmentation mask;

[0154] A space-velocity joint feature model is constructed for the navigation frame sequence. The radial velocity of the ship is fitted using the horizontal footprint center and timestamp of each frame. The fusion weight of each frame is determined by combining the radial velocity with the point cloud coverage of each frame. Frames that are close to the horizontal plane and have high coverage are assigned greater weights to suppress the dilution of the fusion envelope by sparse frames at the ends during multi-frame fusion.

[0155] Specifically, see Figure 6 The fixed bayonet can optionally be equipped with a synchronous camera. Without relying on radar-camera extrinsic parameter calibration or image instance segmentation masks, image target tracking is used to determine the start and end times of a single passage of a vessel from entering to leaving the field of view. This allows for the complete capture of all radar point cloud frames within the time window, forming a navigation frame sequence; among which... This is the moment of first entering the field of view. This is the final moment of departure.

[0156] Further, a space-velocity joint feature model was constructed on the frame sequence, utilizing the horizontal footprint center of the ship in each frame. With timestamp Fitting the radial velocity of the ship Construct joint feature vectors (Used to characterize the variation of radial velocity with heading when a ship passes a fixed radar, and the densest point cloud coverage at near-positive transverse moments), where To remove the rotation of the ship's coordinates, For height, The heading angle is used. This model is used to characterize the variation of the ship's radial velocity with heading, and fusion weights are assigned to each frame accordingly. This allows frames near the positive transverse moment (with low radial velocity and high coverage) to receive greater weight, and these weights are used in multi-frame fusion. By applying weighted or optimized methods, the dilution of the fused envelope by sparse frames at the ends can be effectively suppressed. This space-velocity joint feature model can also be used as a navigation event archive for the identification of abnormal passage (stagnation, backtracking).

[0157] Furthermore, in one embodiment, the method further includes a quality scoring step: based on the multi-frame hull point cloud coverage... Multi-frame freeboard consistency Dimensional anomaly detection results and the results of the consistency verification of drainage volume A comprehensive quality score indicator is generated. When the quality score indicator is lower than a preset quality threshold, a low confidence mark is output or a manual review process is triggered.

[0158] Preferably, in some embodiments, the comprehensive quality score identifier Represented as:

[0159]

[0160] In the formula, All are weighting coefficients; for The corresponding maximum permissible standard deviation, for The corresponding maximum permissible error.

[0161] when When the data falls below a preset threshold, the system retains the original point cloud, bird's-eye view, and intermediate measurements, and marks the navigation operation as low-confidence or transfers it to manual review. This mechanism does not alter the main data chain for size and load condition calculations, but it can reduce the impact of complex water surfaces, obstructions, and abnormal registration data on automated enforcement conclusions.

[0162] See Figure 5 The following comparative examples are used to illustrate the technical advantages of the present invention over prior art solutions and do not constitute a limitation on the scope of protection of the present invention.

[0163] Comparative Example 1: Estimating Vertical Direction Frame by Frame

[0164] Perform RANSAC independently on each frame to find the corresponding... When the global unit normal is calculated, the normal estimation fluctuates with frame jitter, causing the horizontal projection to tilt and resulting in a size distortion of approximately 140m × 76m. This invention estimates a single global u by merging multiple frames, restoring a reasonable magnitude of approximately 48m × 12m on the same data.

[0165] Comparative Example 2: Voxel-based background subtraction

[0166] When the voxel count is not less than When a frame appears, all points within that voxel are deleted ( Choose 0.34, 0.7, or 0.9. (Total number of frames), the hull point in the center of the channel is easily eroded (for example, the ship length is reduced from about 48m to less than 40m). It is still possible to mistakenly select distant tree lines or similar structures. This invention uses grid occupancy to mark the static background and uses a channel mask to limit the search domain without deleting points within the channel.

[0167] Comparative Example 3: Low Freeboard Quantiles and Clustered Algal Subsets

[0168] If we consider the clustering stage The subset of the ship's hull after algae filtering, and with low quantiles The estimated freeboard value for vessels with low freeboard will be constrained to around 1m. This invention estimates the freeboard within the complete point cloud over the horizontal footprint, decoupling it from the algae-filtering threshold.

[0169] Comparative Example 4: Creating a separate background within a single flight sequence

[0170] If the occupancy rate is only calculated based on multiple frames of point cloud data from a single slow-speed passage, the grid cell containing the ship's hull will be marked as static background and removed, resulting in a significantly underestimation of the ship's length. This invention requires the construction of a global background model using single-frame data from multiple ship passages.

[0171] Comparative Example 5: One-dimensional longitudinal cross-correlation registration

[0172] Longitudinal correlation on a smooth hull can easily lead to ambiguity in aperture. This invention achieves multi-frame stitching by rotating each frame's heading and aligning it with the frame center.

[0173] In one embodiment, combined Figure 2 A system for intelligent detection of ship dimensions and loading status based on multi-frame heading correction is provided, for executing the intelligent detection method for ship dimensions and loading status based on multi-frame heading correction. The system includes:

[0174] The first module (point cloud access module) is used to acquire lidar point cloud data collected by fixed monitoring stations and organize single or multi-frame sequences according to navigation events.

[0175] The second module (background and coordinates module) is used to estimate the global vertical direction vector, establish the horizontal projection coordinate system, and construct the static background occupancy model and the channel grid mask.

[0176] The third module (hull separation module) is used to estimate the water surface height and perform geometric clustering segmentation and neighboring fragment merging on the point cloud within the mask-defined search domain;

[0177] The fourth module (size measurement module) is used to measure the size of a single frame using a dual-algorithm optimal orientation strategy, and to estimate the freeboard within the horizontal footprint of the hull using a height histogram;

[0178] The fifth module (Depth and Draft Module) is used to calculate the draft, estimate the displacement, and determine the load condition by combining the global regression model of the depth or the registration database.

[0179] The sixth module (post-processing module) is used to perform multi-frame heading derotation and center alignment fusion on the same navigation sequence, and to perform multi-level verification correction or measurement quality scoring.

[0180] Specific limitations regarding the intelligent detection system for ship dimensions and loading condition based on multi-frame heading correction can be found in the limitations of the intelligent detection method for ship dimensions and loading condition based on multi-frame heading correction mentioned above, and will not be repeated here. Each module in the aforementioned intelligent detection system for ship dimensions and loading condition based on multi-frame heading correction can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0181] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0182] Step 1: Obtain the lidar point cloud data collected by the fixed monitoring station, and estimate the global vertical direction vector based on the multi-frame point cloud to establish a horizontal projection coordinate system;

[0183] Step 2: In the horizontal projection coordinate system, a background occupancy model is constructed by statistically analyzing the frequency of occurrence of spatial grids in point clouds of multiple frames or multiple ship trips, and a channel grid mask is constructed using non-static grids to limit the ship hull search area.

[0184] Step 3: Estimate the water surface height based on the point cloud within the channel grid mask, filter candidate points with a preset margin above the water surface, perform clustering and fragment merging, and obtain the hull point set;

[0185] Step 4: Project the hull point set onto the horizontal plane, and use a dual-algorithm optimal orientation strategy to determine the ship's heading envelope, thereby obtaining the ship's length, width, and height in a single frame;

[0186] Step 5: Within the determined horizontal footprint range of the ship, extract sampling points from the complete point cloud, estimate the main deck height based on the statistical characteristics of the height histogram of the sampling points relative to the water surface height distribution, and calculate the freeboard;

[0187] Step 6: Obtain the ship's molded depth, calculate the draft based on the molded depth and freeboard, and then estimate the displacement and determine the ship's loading status.

[0188] Step 7: For the navigation frame sequence of the same ship, perform derotation and center alignment fusion using the heading angle and envelope geometric center of each frame, and refine the output ship size on the fused point cloud.

[0189] For specific limitations on each step, please refer to the limitations of the intelligent detection method for ship size and loading condition based on multi-frame heading correction mentioned above, which will not be repeated here.

[0190] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0191] Step 1: Obtain the lidar point cloud data collected by the fixed monitoring station, and estimate the global vertical direction vector based on the multi-frame point cloud to establish a horizontal projection coordinate system;

[0192] Step 2: In the horizontal projection coordinate system, a background occupancy model is constructed by statistically analyzing the frequency of occurrence of spatial grids in point clouds of multiple frames or multiple ship trips, and a channel grid mask is constructed using non-static grids to limit the ship hull search area.

[0193] Step 3: Estimate the water surface height based on the point cloud within the channel grid mask, filter candidate points with a preset margin above the water surface, perform clustering and fragment merging, and obtain the hull point set;

[0194] Step 4: Project the hull point set onto the horizontal plane, and use a dual-algorithm optimal orientation strategy to determine the ship's heading envelope, thereby obtaining the ship's length, width, and height in a single frame;

[0195] Step 5: Within the determined horizontal footprint range of the ship, extract sampling points from the complete point cloud, estimate the main deck height based on the statistical characteristics of the height histogram of the sampling points relative to the water surface height distribution, and calculate the freeboard;

[0196] Step 6: Obtain the ship's molded depth, calculate the draft based on the molded depth and freeboard, and then estimate the displacement and determine the ship's loading status.

[0197] Step 7: For the navigation frame sequence of the same ship, perform derotation and center alignment fusion using the heading angle and envelope geometric center of each frame, and refine the output ship size on the fused point cloud.

[0198] For specific limitations on each step, please refer to the limitations of the intelligent detection method for ship size and loading condition based on multi-frame heading correction mentioned above, which will not be repeated here.

[0199] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for intelligent detection of ship dimensions and loading status based on multi-frame heading correction, characterized in that, The method includes the following steps: Step 1: Obtain the lidar point cloud data collected by the fixed monitoring station, and estimate the global vertical direction vector based on the multi-frame point cloud to establish a horizontal projection coordinate system; Step 2: In the horizontal projection coordinate system, a background occupancy model is constructed by statistically analyzing the frequency of occurrence of spatial grids in point clouds of multiple frames or multiple ship trips, and a channel grid mask is constructed using non-static grids to limit the ship hull search area. Step 3: Estimate the water surface height based on the point cloud within the channel grid mask, filter candidate points with a preset margin above the water surface, perform clustering and fragment merging, and obtain the hull point set; Step 4: Project the hull point set onto the horizontal plane, and use a dual-algorithm optimal orientation strategy to determine the ship's heading envelope, thereby obtaining the ship's length, width, and height in a single frame; Step 5: Within the determined horizontal footprint range of the ship, extract sampling points from the complete point cloud, estimate the main deck height based on the statistical characteristics of the height histogram of the sampling points relative to the water surface height distribution, and calculate the freeboard; Step 6: Obtain the ship's molded depth, calculate the draft based on the molded depth and freeboard, and then estimate the displacement and determine the ship's loading status. Step 7: For the navigation frame sequence of the same ship, perform derotation and center alignment fusion using the heading angle and envelope geometric center of each frame, and refine the output ship size on the fused point cloud.

2. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 1, characterized in that, In step 1, the global vertical direction vector is obtained by fitting it on the dominant near-horizontal plane of the scene using a random sampling consensus algorithm, and a constant normal constraint is maintained for all point cloud frames collected by the monitoring station. In step 2, the background occupancy model is specifically a static background occupancy grid model in the horizontal projection coordinate system: The occurrence frequency is defined by the number of frames, that is, the proportion of the number of frames in which the horizontal grid cell appears in the single frame point cloud of multiple shipments is used as the occupancy rate, and the same grid cell is counted only once in the same frame, so that the occupancy rate reflects the probability of the grid cell appearing in the frame sequence rather than the point cloud density, thereby achieving decoupling between the occupancy rate and the point cloud density. Grids with an occupancy rate not lower than a preset threshold are marked as static backgrounds. The channel grid mask is constructed using non-static grids, and no deletion-style background subtraction is performed on the point cloud within the ship hull search area defined by the channel grid mask.

3. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 1, characterized in that, In step 3, the clustering segmentation adopts a density-based spatial clustering algorithm; the fragment merging specifically involves: using qualified clusters whose point counts meet the hard constraints of ship size as hull seed clusters, iteratively merging spatially adjacent fragment clusters with horizontal distances within a preset span to restore the truncated hull line; The hard constraints on ship dimensions include at least: ship length range, ship width range, upper limit of cluster height and lower limit of length-to-width ratio; Step 3, when estimating the water level height, also includes a slow tidal correction: the trend function of water level slowly changing over time is fitted using the water level quantile sequence of the same monitoring station within a preset historical time window, and the water level height is dynamically corrected according to the current timestamp during online measurement to eliminate the influence of intraday river level fluctuations on the water level candidate point screening threshold and freeboard calculation.

4. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 1, characterized in that, The dual-algorithm optimization and orientation strategy described in step 4 is as follows: Simultaneously, the minimum bounding rectangle of the hull point set on the horizontal plane and the principal axis-aligned bounding rectangle based on principal component analysis are obtained; Using the ship's slender geometry prior as the criterion for heading correctness, when the ratio of the aspect ratio of the principal axis-aligned circumscribed rectangle to the aspect ratio of the smallest circumscribed rectangle is greater than a preset selection coefficient, the principal axis-aligned circumscribed rectangle is selected as the final ship heading envelope; otherwise, the smallest circumscribed rectangle is selected. The preset selection coefficient is greater than 1.

5. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 1, characterized in that, In step 6, the method for obtaining the ship's shape depth includes: firstly querying the registration database by ship identification code; if the query fails, then using the measured ship length and width, predicting the shape depth through a single global shape depth regression model. The single global deep regression model includes: When only the ship's length and width are known online, a second-order polynomial model using only the ship's length and width as input is adopted; Given the deadweight tonnage and gross tonnage, a log-linear power-law model is used, which includes the ship's length, width, deadweight tonnage, and gross tonnage as inputs.

6. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 1, characterized in that, The de-rotation and center alignment fusion described in step 7 does not rely on nearest neighbor registration between point clouds. Its specific implementation method is as follows: For the heading angle of each frame in the navigation frame sequence Taking the major axis without direction, the reference heading for a single voyage is obtained by averaging the angled vectors: Frame by frame, the frame center determined by the envelope geometry is used as the translation anchor point for center alignment. Rotation is then performed by combining the heading angle of each frame to avoid aperture degradation issues in longitudinal registration of smooth hulls. Time add Eliminate ambiguity regarding 180° orientation; Before derotation and center alignment fusion, abnormally wide frames are removed using the low quantile of the single-frame ship width distribution in each frame. On the fused point cloud, the length and width of the ship are refined and output along the preset quantile boundary range of the ship's length axis and width axis.

7. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to claim 5, characterized in that, Before outputting dimensions and load conditions in step 7, there are also multi-level verification and anomaly correction steps: The measured ship dimensions are then subjected to size range tests, length-to-beam ratio tests, and regression matching tests of ship beam to ship length, as well as registered Ku Bayesian posterior corrections. The Bayesian posterior correction of the registration library is specifically as follows: taking the conditional distribution of the registration library as the prior, when the measured ship width deviates from the conditional expectation by more than a preset multiple of the standard deviation of the registration library conditions, the Bayesian posterior expectation of the prior and the measured ship width is calculated, and the measured ship width is replaced by the posterior expectation as the final output. After determining the load condition in step 6, the process also includes a consistency check of the discharge volume: The geometric displacement of the ship is calculated and compared with the displacement estimated based on the power law of the registry. When the relative deviation between the two exceeds a preset deviation threshold, the confidence level of the load condition determination is reduced or the ship is marked as suspected of being overloaded.

8. The intelligent detection method for ship size and loading status based on multi-frame heading correction according to any one of claims 1 to 7, characterized in that, The method also includes a single navigation time window delineation step: The ship is visually tracked based on camera images synchronized with the lidar. The start and end times of a single navigation passage from the ship's entry into the field of view to its departure are determined to define the navigation time window. All lidar point cloud frames with corresponding timestamps within the navigation time window are captured to form the navigation frame sequence. The visual tracking does not rely on camera intrinsic and extrinsic parameter calibration and image instance segmentation mask. A space-velocity joint feature model is constructed for the navigation frame sequence. The radial velocity of the ship is fitted using the horizontal footprint center and timestamp of each frame. The fusion weight of each frame is determined by combining the radial velocity with the point cloud coverage of each frame. Frames that are close to the horizontal plane and have high coverage are assigned greater weights to suppress the dilution of the fusion envelope by sparse frames at the ends during multi-frame fusion.

9. A ship size and loading condition intelligent detection system based on multi-frame heading correction, used to execute the method as described in any one of claims 1 to 8, characterized in that, The system includes: The first module is used to acquire lidar point cloud data collected by fixed monitoring stations and organize single or multi-frame sequences according to navigation events. The second module is used to estimate the global vertical direction vector, establish the horizontal projection coordinate system, and construct the static background occupancy model and the channel grid mask. The third module is used to estimate the water level height and perform geometric clustering segmentation and neighbor fragment merging on the point cloud within the mask-defined search domain. The fourth module is used to measure the size of a single frame using a dual-algorithm optimal orientation strategy, and to estimate the freeboard within the horizontal footprint of the hull using a height histogram. The fifth module is used to calculate the draft, estimate the displacement, and determine the loading status by combining the global regression model or the registered database. The sixth module is used to perform multi-frame heading derotation and center alignment fusion on the same navigation sequence, and to perform multi-level verification correction or measurement quality scoring.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images