Unmanned system-oriented multi-source point cloud space unit level dynamic compression method and device

Abstract

The application discloses a multi-source point cloud space unit level dynamic compression method and device for unmanned systems, which comprises the following steps: dividing a three-dimensional space formed by point cloud data collected by multiple point cloud collection sources into multiple space units; calculating space unit state parameters corresponding to each space unit; determining space unit level reserved point numbers according to space unit state parameters of a reference space unit and global budget point numbers; calculating source reliability evaluation values for each point cloud collection source corresponding to point cloud data in each space unit; calculating source level reserved point numbers according to source reliability evaluation values of a reference point cloud collection source in a reference space unit and the space unit level reserved point numbers; determining point cloud data to be reserved according to the source level reserved point numbers and point level priority sorting results of the point cloud data, and merging to obtain compressed point cloud data. The method can improve space coverage uniformity and source reliability under the condition that global budget point numbers are limited.

Description

Technical Field

[0001] This application relates to the fields of 3D point cloud data processing, multi-source collaborative sensing, and point cloud compression technology, specifically to a method and apparatus for dynamic compression of multi-source point cloud spatial units for unmanned systems. Background Technology

[0002] In scenarios involving autonomous driving, mobile robots, drones, roadside sensing units, and multi-robot collaborative perception, multiple point cloud acquisition sources can obtain 3D point cloud data from different locations and perspectives. Multi-source point clouds can alleviate single-viewpoint occlusion and improve the completeness of environmental perception, but they also bring problems such as large point cloud scale, many redundant areas, high communication bandwidth consumption, and high computational overhead for subsequent perception networks.

[0003] Existing point cloud compression methods typically employ uniform compression rates, random sampling, fixed Top-K, or point-level score-based filtering. While these methods can reduce the number of input points, the quality of point clouds from different sources within the same spatial region is inconsistent. Low-reliability sources may consume a large amount of retention budget due to their high number of points; high-risk or sparse spatial regions may have insufficient points retained due to uniform compression rates; and a single high-density source may crowd out other sources with complementary perspectives, thereby reducing the coverage balance of the compressed point cloud.

[0004] In actual deployments, there are time delays, pose errors, differences in ranging range, differences in viewpoint coverage, differences in local point density, and differences in sensor status among point cloud acquisition sources. If importance is only calculated at the point level and directly sorted and retained, it is difficult to explicitly constrain how many points should be retained in each spatial unit, and it is also difficult to constrain how many point budgets each point cloud acquisition source should occupy in the same spatial unit. Under the condition of limited global point budget, it is difficult to guarantee the reliability of point cloud data sources. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for dynamic compression of multi-source point cloud spatial units for unmanned systems, the method comprising: The three-dimensional space formed by point cloud data collected from multiple point cloud acquisition sources is divided into multiple spatial units; the point cloud acquisition sources are deployed on an unmanned system. Based on the information of the point cloud data and the source metadata of the point cloud acquisition source, calculate the spatial unit state parameters corresponding to each spatial unit; the spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. The number of points to be retained at the spatial unit level for the reference spatial unit is determined based on the spatial unit state parameters and the global budget points. The number of points to be retained at the spatial unit level represents the number of point clouds that the reference spatial unit is allowed to retain after compression. The reference spatial unit can be any one of the spatial units. For each point cloud acquisition source corresponding to the point cloud data in each spatial unit, a source reliability assessment value is calculated based on the corresponding source metadata; the source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. Based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of points retained at the spatial unit level of the reference spatial unit, the number of points retained at the source level is calculated; the number of points retained at the source level represents the number of point clouds acquired by the reference point cloud acquisition source that is allowed to be retained in the reference spatial unit; the reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit. Based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, the point cloud data to be retained is determined and merged to obtain compressed point cloud data.

[0006] Optionally, the number of reserved points at the spatial unit level of the reference spatial unit is determined based on the spatial unit state parameters and the global budget points. This includes: calculating the retention weight of the reference spatial unit based on the spatial unit state parameters; calculating the ratio of the retention weight of the reference spatial unit to the sum of the retention weights of all spatial units, and multiplying it by the global budget points to obtain the initial number of reserved points at the spatial unit level of the reference spatial unit; and truncating the initial number of reserved points at the spatial unit level using the preset minimum number of reserved points and the preset maximum number of reserved points of the spatial unit to obtain the number of reserved points at the spatial unit level of the reference spatial unit.

[0007] Optionally, the spatial unit state parameters further include point density, target candidate strength, geometric abruptness, and risk state quantity; the calculation of the retention weight of the reference spatial unit based on the spatial unit state parameters of the reference spatial unit includes: normalizing the number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity of the reference spatial unit respectively to obtain the normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity; and weighting and summing the normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity to obtain the retention weight of the reference spatial unit.

[0008] Optionally, the method further includes: increasing the retention weight of the reference spatial cell when the target candidate intensity of the reference spatial cell is greater than the target intensity threshold, or the geometric abruptness of the reference spatial cell is greater than the geometric abruptness threshold; and decreasing the retention weight of the reference spatial cell when the reference spatial cell contains only a dense background and the coverage overlap is greater than the coverage overlap threshold.

[0009] Optionally, the number of source-level retained points is calculated based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit. This includes: calculating the source allocation weight of the reference point cloud acquisition source in the reference spatial unit based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit, the view complementarity factor, and the effective state value of the point cloud data acquired by the reference point cloud acquisition source; the view complementarity factor is determined based on the differences in observation direction, incident angle, visible area, or occlusion relationship of different reference point cloud acquisition sources to the same reference spatial unit; the ratio of the source allocation weight to the sum of the source allocation weights of each point cloud acquisition source in the reference spatial unit is calculated and multiplied by the number of retained points at the spatial unit level of the reference spatial unit to obtain the number of source-level retained points corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0010] Optionally, the point cloud data to be retained is determined based on the number of retained points at the source level and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source. This includes: correcting the number of retained points at the source level and the number of retained points at the spatial unit level based on budget allocation constraints to obtain the corrected number of retained points at the source level; during the correction process, the sum of the number of retained points at each source level does not exceed the number of retained points at the spatial unit level; and determining the point cloud data to be retained based on the corrected number of retained points at the source level and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source.

[0011] Optionally, the budget allocation constraints include, sequentially, a reliability lower limit constraint, a source proportion constraint, a coverage lower limit constraint, a budget upper limit constraint, and a total budget constraint; the correction of the number of source-level reserved points and the number of spatial unit-level reserved points based on the budget allocation constraints includes: First, a reliability lower limit constraint is applied: if the source reliability assessment value is less than the reliability lower limit threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is set to zero; if the source reliability assessment value is greater than or equal to the reliability lower limit threshold and less than the source normal allocation threshold, and the view complementarity factor of the reference point cloud acquisition source in the reference spatial unit is greater than the complementarity threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is corrected to the first source-level reserved point number; the first source-level reserved point number is the preset minimum source-level reserved point number. After applying the reliability lower limit constraint, the source proportion constraint is applied: if the number of first source-level reserved points is greater than the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit, the number of first source-level reserved points is adjusted to the number of second source-level reserved points; the number of second source-level reserved points is the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit. After implementing the source proportion constraint, the coverage lower limit constraint is implemented: the effective point cloud acquisition sources in the reference spatial unit are determined to have a target candidate intensity greater than the target intensity threshold, a risk state quantity greater than the risk threshold, or a view complementarity factor greater than the complementarity threshold, and the number of source-level retained points of the effective point cloud acquisition sources is increased so that the sum of the source-level retained points of each point cloud acquisition source in the reference spatial unit reaches the preset minimum number of retained points of the spatial unit or there are no effective point clouds available for retention; After applying the lower limit constraint on coverage, apply the upper limit constraint on budget: if the sum of the source-level reserved points of each point cloud acquisition source in the reference spatial unit is greater than the preset maximum reserved point number of the spatial unit, reduce the source-level reserved point number of each point cloud acquisition source according to the source allocation weight or point-level priority, and determine the sum of the corrected source-level reserved points as the final spatial unit-level reserved point number of the reference spatial unit; the final spatial unit-level reserved point number is less than or equal to the preset maximum reserved point number of the spatial unit. After the budget cap constraint is executed, the overall budget constraint is executed: if the sum of the final spatial unit-level reserved points of each spatial unit exceeds the global budget points, the corresponding reduction priority is calculated based on the risk state quantity, source average reliability, and coverage overlap of each specific spatial unit, and the spatial unit-level reserved points of the specific spatial unit are reduced based on each reduction priority until the sum of the final spatial unit-level reserved points of each spatial unit satisfies the global budget points; whereby, a specific spatial unit is a spatial unit whose risk state quantity is less than the risk state threshold, whose source average reliability is less than the spatial unit reliability threshold, and whose coverage overlap is greater than the redundancy threshold.

[0012] Optionally, the point cloud data to be retained is determined based on the corrected source-level retained points and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source. This includes: selecting candidate point cloud data from the reference spatial unit with a number not exceeding the corrected source-level retained points of the reference point cloud acquisition source based on the point-level priority ranking results of the point cloud data collected by the reference point cloud acquisition source; and using the candidate point cloud data as the point cloud data to be retained corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0013] Optionally, the point-level priority ranking result is determined based on at least one of the following: distance from the point cloud data to the sensor corresponding to the point cloud acquisition source, reflection intensity, local point density, local curvature, and target candidate intensity.

[0014] This application also provides a multi-source point cloud spatial unit-level dynamic compression device for unmanned systems, comprising: The spatial partitioning module is used to divide the three-dimensional space composed of point cloud data collected by multiple point cloud acquisition sources into multiple spatial units; the point cloud acquisition sources are deployed on the unmanned system; The spatial state parameter calculation module is used to calculate the spatial unit state parameters corresponding to each spatial unit based on the information of the point cloud data and the source metadata of the point cloud acquisition source; the spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. The spatial unit budget allocation module is used to determine the number of spatial unit-level retained points of the reference spatial unit based on the spatial unit state parameters and the global budget points. The number of spatial unit-level retained points represents the number of point clouds that the reference spatial unit is allowed to retain after compression. The reference spatial unit can be any one of the spatial units. The source reliability assessment value calculation module is used to calculate the source reliability assessment value for each point cloud acquisition source corresponding to the point cloud data in each spatial unit, based on the corresponding source metadata. The source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. The source-level budget allocation module is used to calculate the number of source-level reserved points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of spatial unit-level reserved points of the reference spatial unit. The number of source-level reserved points represents the number of point clouds acquired by the reference point cloud acquisition source that is allowed to be retained in the reference spatial unit. The reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit. The point cloud selection and compression module is used to determine the point cloud data to be retained based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, and to merge them to obtain compressed point cloud data.

[0015] In this embodiment, the three-dimensional space composed of point cloud data collected by multiple point cloud acquisition sources is divided into multiple spatial units. Based on the information of the point cloud data and the source metadata of the point cloud acquisition sources, the spatial unit state parameters corresponding to each spatial unit are calculated. Based on the spatial unit state parameters of the reference spatial unit and the global budget points, the number of spatial unit-level retained points for the reference spatial unit is determined. For each point cloud acquisition source corresponding to the point cloud data in each spatial unit, a source reliability assessment value is calculated based on the corresponding source metadata. Based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of spatial unit-level retained points for the reference spatial unit, the number of source-level retained points is calculated. Based on the number of source-level retained points and the point-level priority ranking result of the point cloud data collected by each point cloud acquisition source, the point cloud data to be retained is determined and merged to obtain compressed point cloud data.

[0016] Thus, based on the information of point cloud data and the source metadata of point cloud acquisition sources, the spatial unit state parameters of spatial units are calculated, elevating the assessment of each spatial unit from a simple statistical count of the number of point cloud data points to a comprehensive evaluation of its retention value. The spatial unit state parameters, including the number of point cloud acquisition sources and coverage overlap, characterize the multi-source coverage and degree of repeated observation within the same spatial unit, providing a basis for determining the number of retained points at the spatial unit level through multi-source collaborative sensing. Based on the spatial unit state parameters and the global budget points of the reference spatial unit, the number of retained points at the spatial unit level is determined, enabling differentiated allocation of global budget points based on comprehensive value. This allows valuable or high-risk areas to receive more point allocations, while highly redundant background areas receive fewer, thus optimizing the allocation of limited resources in the spatial dimension. Furthermore, the source reliability assessment value is calculated based on the source metadata of the point cloud acquisition sources, enabling a quantitative assessment of the data credibility of each point cloud acquisition source. This allows the system to identify inferior sources with time delays, pose drift, or poor observation conditions, providing a basis for subsequent source... The allocation of retention points at the source level provides a quality basis; the source reliability assessment value is determined at least based on the time freshness of point cloud data acquisition and the pose confidence of the point cloud acquisition source, which can simultaneously reflect the impact of time delay and pose error in multi-source point clouds. This can avoid low-freshness or low pose confidence sources occupying too much budget in the source-level retention point allocation, and improve the multi-source coverage balance; the source-level retention point number is calculated based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the spatial unit-level retention point number of the reference spatial unit. This realizes the further splitting of the spatial unit-level retention budget into dedicated quotas for each acquisition source based on the source reliability assessment value, so that unreliable sources cannot occupy the budget, and improve the reliability of point cloud data sources; the point cloud data to be retained is determined based on the source-level retention point number and the point-level priority ranking result of the point cloud data acquired by each point cloud acquisition source. That is, only after the spatial unit-level retention point number and the source-level retention quota are determined, point clouds are selected in each spatial unit and each point cloud acquisition source according to the point-level priority ranking result, thereby achieving precise control of the retention point number of different spatial regions and different sources. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a method for dynamic compression of multi-source point cloud spatial units for unmanned systems, provided in an embodiment of this application; Figure 2This is a schematic diagram of the structure of a multi-source point cloud spatial unit-level dynamic compression device for unmanned systems, provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] It should be noted that the terms "first" and "second" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0022] This application provides a multi-source point cloud spatial cell-level dynamic compression method for unmanned systems. This method can be applied to collaborative sensing and communication scenarios and implemented through processing devices (such as local devices or edge nodes). Figure 1 The diagram shown is a flowchart illustrating a multi-source point cloud spatial cell-level dynamic compression method for unmanned systems provided in this application. The method includes: S101. Divide the three-dimensional space composed of point cloud data collected by multiple point cloud acquisition sources into multiple spatial units.

[0023] Among them, point cloud acquisition sources can be deployed on unmanned systems. Unmanned systems refer to intelligent systems with autonomous movement as their core, including but not limited to autonomous vehicles, mobile robots, drones, multi-robot systems, and roadside collaborative perception infrastructure serving these systems.

[0024] In some embodiments, point cloud data collected by multiple point cloud acquisition sources can be acquired, as well as source metadata of each point cloud acquisition source. The source metadata may include acquisition source number, acquisition timestamp, sensor pose, external parameter calibration information, sensor status, etc.

[0025] In some embodiments, methods such as voxels, cylinders, regular meshes, spherical meshes, or irregular spatial partitioning can be used to divide the unit space of point cloud data collected by multiple acquisition sources into multiple spatial units, and record the spatial unit to which each point cloud data belongs and the corresponding point cloud acquisition source, thereby forming a correspondence between spatial units and sources.

[0026] S102. Based on the information of the point cloud data and the source metadata of the point cloud acquisition source, calculate the spatial unit state parameters corresponding to each spatial unit.

[0027] The spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. Point cloud data information can include the spatial coordinates of the point cloud data and the corresponding point cloud acquisition sources. The number of point cloud acquisition sources indicates how many point cloud acquisition sources cover the spatial unit and can be determined based on the point cloud acquisition sources corresponding to the point cloud data. Coverage overlap indicates whether the distribution of point cloud data acquired by different point cloud acquisition sources within the spatial unit is duplicated and can be determined based on the coverage number of different point cloud acquisition sources within the same spatial unit, the spatial coordinates of the point cloud data, the intersection-over-union ratio of the point cloud distribution, or the proportion of overlapping spatial coverage areas.

[0028] In some embodiments, the spatial cell state parameters further include at least one of point density, target candidate strength, geometric abruptness, and risk state quantity.

[0029] In some embodiments, point density is used to characterize the spatial density of point cloud points within a spatial cell. Spatial cell Point density Based on the number of point cloud points contained within a spatial unit With the volume of this space unit The ratio is determined, that is .

[0030] In some embodiments, the target candidate strength is used to characterize the probability that a spatial cell contains a target of interest. In one embodiment, the maximum or average target confidence score output by the target candidate detector for each point cloud point or (point cloud cluster) within the reference spatial cell can be determined as the target candidate strength. In another embodiment, the target candidate strength can also be determined by one of the following: the spatial overlap between the target trajectory in the historical frame and the spatial cell, the foreground probability, and a preset risk area identifier. Alternatively, the maximum value among the spatial overlap between the target trajectory in the historical frame and the reference spatial cell, the foreground probability, and the preset risk area identifier can be used as the target candidate strength.

[0031] In some embodiments, geometric abruptness is used to characterize the degree of geometric discontinuity of the point cloud surface within a spatial cell. In one embodiment, for each point cloud point within a spatial cell, a local neighborhood containing that point and its neighboring points is constructed (the local neighborhood is determined by the set of point cloud points within the current voxel), and the covariance matrix of the coordinates of the local neighborhood points, as well as the eigenvalues ​​of the covariance matrix, are calculated. These three eigenvalues ​​describe the dispersion of the point and its neighbors in three directions in three-dimensional space, satisfying the following conditions: , Corresponding to the principal direction (representing the direction of greatest dispersion, usually extending along the object's surface), if the point cloud is approximately on a plane, Pointing in the direction where the point cloud spreads most widely within the plane (e.g., along the length of a highway); The corresponding secondary direction describes the degree of dispersion of the point cloud in the second dimension perpendicular to the principal direction. The corresponding normal direction (representing the direction of maximum concentration) describes the degree of dispersion (i.e., thickness) of the point cloud perpendicular to the fitting plane; the local curvature of point p in the point cloud. ,in To prevent the use of a pre-defined positive number with a denominator of zero, the average local curvature of each point cloud point within the spatial cell is determined as the geometric abrupt change degree.

[0032] In some embodiments, the risk state quantity is used to characterize the degree of risk of collision, occlusion, or sensing failure within a spatial cell. In one embodiment, the risk state quantity can be determined based on whether the spatial cell is located at an intersection, narrow passage, occlusion boundary, or a pre-defined high-risk area, and at least one of the relative distance, relative velocity, and expected collision time of dynamic targets within the spatial cell. Each state parameter is linearly normalized based on the sensor range and the upper and lower limits of the corresponding state parameters in the validation data, and the output is limited to the range [0, 1] by batch normalization and the Sigmoid function.

[0033] S103. Determine the number of space cell-level reserved points for the reference space cell based on the space cell state parameters and global budget points of the reference space cell.

[0034] In some embodiments, the reference spatial unit can be any one of the spatial units. Based on the spatial unit state parameters of the reference spatial unit, the global budget points can be allocated differentially to obtain the spatial unit-level retained points of the reference spatial unit. The spatial unit-level retained points represent the number of point clouds that the reference spatial unit is allowed to retain after compression.

[0035] In some embodiments, determining the number of reserved points at the spatial unit level for a reference spatial unit based on the spatial unit state parameters and the global budget points includes: calculating the reservation weight of the reference spatial unit based on the spatial unit state parameters; calculating the ratio of the reservation weight of the reference spatial unit to the sum of the reservation weights of all spatial units, and multiplying it by the global budget points to obtain the initial number of reserved points at the spatial unit level for the reference spatial unit; and truncating the initial number of reserved points at the spatial unit level using the preset minimum number of reserved points and the preset maximum number of reserved points for the spatial unit to obtain the number of reserved points at the spatial unit level for the reference spatial unit.

[0036] In some embodiments, the global budget points can be dynamically adjusted based on the current communication bandwidth, the target processing network input limit, edge computing load, or latency constraints. When bandwidth decreases, the global budget points are reduced. When communication bandwidth is low, the global budget points can be reduced; when a high-risk spatial region or critical target region is detected, the budget weight of the corresponding spatial unit can be increased. In this way, the compression process can dynamically change with communication bandwidth and task risk, without requiring a fixed compression rate.

[0037] In some embodiments, calculating the retention weight of a reference spatial unit based on its spatial unit state parameters includes: normalizing the number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity of the reference spatial unit, respectively, to obtain normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity; and weighting and summing the normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity to obtain the retention weight of the reference spatial unit. The normalization of the number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity of the reference spatial unit can be achieved by mapping these parameters to the range [0, 1].

[0038] For example, reference space unit Retention weight It can be calculated using formula (1): (1); in, , and These represent the normalized point cloud data collection source quantity, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity, respectively. and These represent the weights corresponding to the number of point cloud acquisition sources, coverage overlap, point density, target candidate intensity, geometric abruptness, and risk state quantity, respectively, after normalization. Furthermore, the direction of the point density term should be determined based on its actual meaning of representing effective information or background redundancy; this embodiment adopts a positive direction.

[0039] In some embodiments, the weights corresponding to the number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric mutation degree, and risk state quantity after normalization are all non-negative numbers. Each weight can be generated based on the data-driven and task-oriented adaptive learning mechanism of the TSK fuzzy inference system. Specifically, the TSK system first uses Gaussian membership functions to convert state parameters (such as geometric abruptness, risk state quantity, etc.) from precise values ​​to fuzzy membership degrees. The center parameter and scale parameter of the Gaussian function are initialized based on the statistical characteristics of the training data (such as mean and standard deviation). Subsequently, the TSK system constructs multiple "if-then" fuzzy rules, each rule corresponding to a linear consequent equation. The coefficients in the consequent equation are equivalent to the weights (learnable variables) in formula (1). Finally, the output of the TSK system (i.e., the retained weights of spatial units) is applied to the downstream 3D target detection task. The detection loss error is propagated back layer by layer through the backpropagation algorithm. The coefficients in the consequent equations of each fuzzy rule are adjusted end-to-end. After a large amount of training data iteration, the system can automatically evolve a set of optimal weight combinations, thereby achieving adaptive budget allocation according to the data characteristics and the needs of the perception task.

[0040] In some embodiments, the initial number of reserved points at the spatial unit level is truncated using a preset minimum number of reserved points and a preset maximum number of reserved points. This can be achieved by determining the preset minimum number of reserved points as the number of reserved points at the spatial unit level for the reference spatial unit if the initial number of reserved points at the spatial unit level is less than the preset minimum number of reserved points, and by determining the preset maximum number of reserved points as the number of reserved points at the spatial unit level for the reference spatial unit if the initial number of reserved points at the spatial unit level is greater than the preset maximum number of reserved points.

[0041] In some embodiments, the retention weight of a reference spatial cell can increase with the increase of the target candidate strength, geometric abruptness, and risk state quantity of the reference spatial cell, and decrease with the increase of coverage overlap or background redundancy. The retention weight of the reference spatial cell is increased when the target candidate strength of the reference spatial cell is greater than a target strength threshold, or when the geometric abruptness of the reference spatial cell is greater than a geometric abruptness threshold; the retention weight of the reference spatial cell is decreased when the reference spatial cell contains only a dense background and the coverage overlap is greater than a coverage overlap threshold.

[0042] In some embodiments, when the target candidate intensity of a reference spatial cell is greater than a target intensity threshold, or the geometric abruptness is greater than a geometric abruptness threshold, the retention weight of the reference spatial cell can be increased by a preset increment or multiplied by a first adjustment coefficient greater than 1. In other embodiments, when a reference spatial cell simultaneously satisfies the following conditions: target candidate intensity is lower than a first background threshold, geometric abruptness is lower than a second background threshold, risk state quantity is lower than a third background threshold, and point density is higher than a point density threshold, the spatial cell is determined as a dense background spatial cell. Furthermore, when the coverage overlap of the dense background spatial cell is greater than a coverage overlap threshold, its retention weight is reduced by a preset decrease or multiplied by a second adjustment coefficient less than 1. It should be noted that the various thresholds mentioned above (target intensity threshold, geometric abruptness threshold, first background threshold, second background threshold, third background threshold, and point density threshold) do not have fixed values ​​across scenes, but are adaptively determined through the center parameter and scale parameter of the Gaussian membership function corresponding to each state parameter. Specifically, for each threshold corresponding to a state parameter, the numerical distribution of the state parameter in the pre-training dataset is collected, and its mean and standard deviation are calculated. These are then used to initialize the center parameter and scale parameter of the corresponding Gaussian membership function, thereby determining the threshold corresponding to each state parameter (when Gaussian membership is used as the criterion, the threshold is equal to a linear combination of the center parameter and the scale parameter). In this way, each threshold can be synchronously and adaptively updated as the scene changes, without the need for manual parameter tuning. The first and second adjustment coefficients are determined by validation data or end-to-end training. The preset increment and preset decrement are configurable constants that can be preset by engineers based on system requirements, experimental experience, or offline calibration.

[0043] In some embodiments, spatial unit The corresponding number of spatial unit level reserved points It can be calculated using formula (2): (2); in, Represents the total budget points. Representing spatial units The retention weight, Representing spatial units The retention weight, This represents the sum of the retention weights of each spatial unit. and These represent the preset minimum and maximum number of reserved points for a spatial unit, respectively. This indicates rounding to the nearest integer. This indicates that the number of reserved points at the spatial unit level is limited to [a certain number]. and Between. When the truncation process causes the sum of the retained points at each spatial unit level to exceed the global budget points, the total budget constraint described later is executed.

[0044] S104. For each point cloud acquisition source corresponding to the point cloud data in each spatial unit, calculate the source reliability assessment value based on the corresponding source metadata.

[0045] The source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. Time freshness can be determined based on the difference between the current processing time and the point cloud acquisition timestamp; pose confidence can be determined based on the positioning covariance (positioning drift probability given by GPS / INS), calibration error (small deflection of sensor mounting angle), or registration residual (matching error after point cloud stitching).

[0046] In some embodiments, the source reliability assessment value is further determined based on at least one of the following: observation distance of the point cloud acquisition source to the spatial cell, spectral coverage, point cloud density, sensor status, and historical consistency. The observation distance can be determined based on the Euclidean distance between the spatial location of the point cloud acquisition source and the center of the spatial cell. Spectral coverage characterizes the effective visibility of the point cloud acquisition source relative to the spatial cell and is determined by any one of the following: incident angle, occlusion relationship, and complementary spectral degree. The incident angle is determined based on the angle between the direction vector from the point cloud acquisition source to the center of the spatial cell and the normal vector of the point cloud surface within the spatial cell. The occlusion relationship is determined by emitting a ray from the point cloud acquisition source location to the center of the spatial cell and detecting whether there are occlusion points along the path. The complementary spectral degree is determined based on the angular differences between the direction vectors of multiple point cloud acquisition sources to the same center of the spatial cell. Historical consistency characterizes the stability between the current frame point cloud and historical observation data and is determined by any one of the following: cross-frame registration residual, consistency with the fusion result, or continuous frame target detection stability. Among them, the cross-frame registration residual is determined by the root mean square error after registering the current frame point cloud with the historical frame point cloud; the consistency with the fusion result is determined by comparing the deviation between the current frame point cloud and the multi-frame fused point cloud; the stability of continuous frame target detection is determined by comparing the consistency of the detection box position, size and orientation of the same target in continuous frames.

[0047] In some embodiments, time freshness, pose confidence, observation distance, view coverage, point cloud density, sensor state, and historical consistency can be normalized to obtain normalized reliability factors (including time freshness factor, pose confidence factor, distance attenuation factor, view coverage factor, sensor state factor, and historical consistency factor). These normalized reliability factors are then multiplied or weighted and summed to obtain the source reliability assessment value. During the weighted summation of the reliability factors, the weights of each factor can be determined by searching the validation dataset, ensuring that after incorporating preset time delays, pose errors, distance variations, or occlusion disturbances, the source reliability assessment value has a preset correlation with the corresponding source's point cloud registration accuracy, occlusion area coverage, or downstream sensing performance. For example, TSK fuzzy inference combination can be used, where each reliability factor is converted to a membership degree via a Gaussian membership function, and then weighted fusion is performed using trainable consequent parameters. During the weight verification process, the added pose perturbation is generated by sampling through a Gaussian distribution, where the standard deviation of the position error is in the range of 0 to 0.5 meters and the standard deviation of the heading error is in the range of 0° to 0.5°, in order to cover a variety of cases from no error to significant error, and to ensure the robustness and generalization ability of the weights under different pose accuracy conditions.

[0048] In some embodiments, when the time difference between the point cloud acquisition time and the current processing time is greater than a time delay threshold, the pose confidence level is less than a pose confidence level threshold, or the distance from the point cloud acquisition source to the center of the reference spatial cell is greater than a distance threshold, the source reliability assessment value of the point cloud acquisition source is reduced. When the view coverage of the point cloud acquisition source to the reference spatial cell is greater than a view coverage threshold, or the view complementarity factor between the point cloud acquisition source and other point cloud acquisition sources is greater than a complementarity threshold, the source reliability assessment value of the point cloud acquisition source is increased.

[0049] In the NuScenes embodiment, the distance threshold is set to 70 meters; the verification range for position error is 0~0.5 meters, and the verification range for heading error is 0°~0.5°; the time delay threshold is set based on the system's synchronization accuracy. For example, if the time synchronization accuracy is ±20ms, the threshold can be set slightly larger than this value (e.g., 30ms) to ensure that data within the normal synchronization range is not misjudged as outdated; the pose confidence threshold is set based on the statistical distribution of positioning errors. For example, if 95% of the positioning errors are within 0.3 meters, the threshold can be set near this statistical value, and exceeding this range is considered as unreliable positioning; the field of view coverage threshold is set based on the sensor's field of view coverage capability and the proportion of the visible area. The wider the field of view coverage, the lower the threshold can be, and vice versa; the complementarity threshold is calibrated by the downstream perception performance (e.g., target detection accuracy) on the validation set. By testing the detection recall rate under different complementary factor values, the critical value that brings the best performance improvement is selected as the complementarity threshold. During the threshold verification process, the introduced position error is sampled between 0 and 0.5 meters, and the heading error is sampled between 0 and 0.5 degrees to simulate the system performance under different pose accuracy conditions, ensuring that each threshold has robustness and generalization ability.

[0050] In some embodiments, the average nearest neighbor distance is used as the registration residual between the point cloud acquisition source and other point cloud acquisition sources or fused point clouds. When the average registration residual of the point cloud acquisition source in L consecutive frames is greater than the deviation threshold, or the spatial intersection-union ratio of the target detection result and the fusion result in L consecutive frames is less than the consistency threshold, it is determined that the point cloud acquisition source has a large deviation and its source reliability assessment value is reduced. When the point cloud acquisition source provides more than a preset number of effective point clouds for the target boundary or occluded area in L consecutive frames and the corresponding registration residual is not greater than the deviation threshold, it is determined that it provides stable supplementary information and improves the source reliability assessment value. L uses a 13-frame sequence containing the current keyframe and 6 frames before and after it as a statistical window. The deviation threshold and consistency threshold are determined based on the verification data. In the verification data, the source reliability judgment results under different threshold combinations are compared with the manually labeled sensor positioning error and detection deviation. The threshold combination that makes the reliability assessment value most correlated with the actual sensor positioning accuracy is selected as the final value of the deviation threshold and consistency threshold.

[0051] S105. Calculate the number of source-level retained points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit.

[0052] Among them, the number of source-level retained points represents the number of point clouds collected by the reference point cloud acquisition source that is allowed to be retained in the reference spatial unit. The reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit.

[0053] In some embodiments, after determining the number of spatial cell-level reserved points for the reference spatial cell, the number of spatial cell-level reserved points can be allocated as the number of source-level reserved points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial cell.

[0054] In some embodiments, the source reliability assessment values ​​of each point cloud acquisition source within the same spatial unit can be normalized to obtain the source allocation weight of the reference point cloud acquisition source in the reference spatial unit. The initial source-level reserved point number of the reference point cloud acquisition source is then determined based on the spatial unit-level reserved point number and the source allocation weight. Further, if the reference point cloud acquisition source has no valid point cloud within the spatial unit or its source reliability assessment value is lower than a preset reliability assessment value, the initial source-level reserved point number of the reference point cloud acquisition source can be set to zero or limited to a preset minimum source-level reserved point number, thereby obtaining the source-level reserved point number of the reference point cloud acquisition source.

[0055] In some embodiments, the number of source-level retained points is calculated based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit, including: Based on the source reliability assessment value, view complementarity factor, and effective state value of the point cloud data acquired by the reference point cloud acquisition sources in the reference spatial unit, the source allocation weight of the reference point cloud acquisition sources is calculated. The ratio of the source allocation weight to the sum of the source allocation weights of all point cloud acquisition sources in the reference spatial unit is calculated, and multiplied by the number of retained points at the spatial unit level of the reference spatial unit to obtain the number of retained points at the source level corresponding to the reference point cloud acquisition source in the reference spatial unit. Among them, the view complementarity factor is determined based on the differences in observation direction, incident angle, visible area, or occlusion relationship of different reference point cloud acquisition sources to the reference spatial unit.

[0056] In some embodiments, spatial unit Midpoint cloud data source Source allocation weight according to Calculation. Among them, This indicates the source reliability assessment value. Indicates complementary perspective factors. Indicates a valid state value; when the point cloud acquisition source has a valid point cloud within the spatial unit. ,otherwise .

[0057] In some embodiments, spatial unit Midpoint cloud data source Source-level retention points Calculate according to formula (3): (3); in, Representing spatial units The set of point cloud acquisition sources containing valid point clouds; This indicates rounding down. Since rounding down may result in the sum of the number of retained points at each source level being less than the number of retained points at the spatial unit level, the remaining points can be allocated from largest to smallest according to the decimal part of the corresponding value of each point cloud acquisition source before rounding, until the sum of the number of retained points at each source level reaches the number of retained points at the spatial unit level or there are no valid point clouds available for retention. Representing spatial units Reference point cloud acquisition source The source reliability assessment value.

[0058] Understandably, by calculating the source reliability assessment value, it is possible to avoid low-reliability sources or high-redundancy sources occupying too much point budget, while retaining source point clouds with complementary perspectives and high reliability. This mechanism differs from the approach of unifying Top-K only within a spatial cell, because it explicitly limits the point quota for each source within the same spatial cell.

[0059] S106. Based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, determine the point cloud data to be retained and merge them to obtain compressed point cloud data.

[0060] The point-level priority ranking result indicates the order in which point cloud data collected by different point cloud acquisition sources are retained; the earlier the ranking, the higher the priority of retention. The point-level priority ranking results may differ for point cloud data collected by different point cloud acquisition sources.

[0061] In some embodiments, for spatial units Internally sourced from point cloud Point cloud data collected Calculate the normalized distance from each point in the point cloud to the sensor. Normalized reflectance Normalized local point density Normalized local curvature and normalized target candidate strength Point cloud Point-level priority score Calculate according to formula (4): (4); in All weights are non-negative and are part of the TSK consequent parameters, learned by backpropagation of the downstream 3D target detection loss.

[0062] In some embodiments, according to Point cloud points are sorted from high to low. When scores are the same, they are sorted according to distance to sensor, acquisition time and point identifier in turn to obtain point-level priority sorting results. In this process, each feature (distance to sensor, acquisition time and point identifier) ​​is first batch normalized and then mapped to (0, 1) using the Sigmoid function.

[0063] In other embodiments, the point-level priority ranking result is obtained by the TSK fuzzy inference system. For each point cloud point, the input attributes include the distance from the point to the scanner, the angle of the point relative to the scanner, the reflection intensity of the point, and the distance from the point to the voxel center. Local region attributes include the point density, mean, median, and variance of the height within the voxel where the point is located, the mean, median, and variance of the reflection intensity, the curvature of the point, and the SHOT histogram feature. Gaussian membership functions are set for each input attribute to convert each attribute value into a corresponding membership degree. The system establishes multiple fuzzy rules, each rule containing an antecedent (fuzzy conditional judgment for each input attribute) and a consequent (the output linear function corresponding to the rule). For the current input point, the degree to which it satisfies the antecedent of each rule (i.e., the trigger strength) is calculated. After normalizing the trigger strength of all rules, the weighted summation of the consequent outputs of each rule is performed to obtain the point-level saliency score of the point. This score is the point-level priority output by the TSK fuzzy inference system, which is used to guide the point selection process during compression. During training, the saliency score is multiplied element-wise with the point-by-point features extracted by the one-dimensional convolutional network and participates in the loss calculation of downstream tasks. Backpropagation optimizes the Gaussian membership function parameters and consequent parameters of each rule in the TSK system, enabling the saliency score to more accurately reflect the retention value of the points. During inference, the convolutional network is no longer needed; the pre-trained TSK system is used directly to calculate the saliency score for each point, and within each voxel, points are sorted from highest to lowest according to this score, thus obtaining the point-level priority ranking result.

[0064] In some embodiments, determining the point cloud data to be retained based on the number of source-level retained points and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source includes: correcting the number of source-level retained points and the number of spatial unit-level retained points based on budget allocation constraints to obtain a corrected number of source-level retained points; and determining the point cloud data to be retained based on the corrected number of source-level retained points and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source.

[0065] In some embodiments, during the process of correcting the number of source-level reserved points and the number of spatial unit-level reserved points based on budget allocation constraints, both the number of source-level reserved points and the number of spatial unit-level reserved points may be adjusted, but the sum of the final number of source-level reserved points obtained through the correction shall not exceed the number of spatial unit-level reserved points, that is, the number of spatial unit-level reserved points constrains the sum of the number of source-level reserved points.

[0066] In some embodiments, budget allocation constraints include a reliability lower bound constraint, a source proportion constraint, a coverage lower bound constraint, a budget upper bound constraint, and a total budget constraint, executed sequentially; the source-level reserved point number and the spatial unit-level reserved point number are adjusted based on the budget allocation constraints, including: First, a reliability lower limit constraint is applied: if the source reliability assessment value is less than the reliability lower limit threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is set to zero according to the reliability lower limit constraint; if the source reliability assessment value is greater than or equal to the reliability lower limit threshold and less than the source normal allocation threshold, and the view complementarity factor of the reference point cloud acquisition source in the reference spatial unit is greater than the complementarity threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is corrected to the first source-level reserved point number according to the reliability lower limit constraint; the first source-level reserved point number is the preset minimum source-level reserved point number. After applying the reliability lower limit constraint, the source proportion constraint is applied: if the number of first source-level reserved points is greater than the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit, the number of first source-level reserved points is adjusted to the number of second source-level reserved points; the number of second source-level reserved points is the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit. After implementing the source proportion constraint, the coverage lower limit constraint is implemented: the effective point cloud acquisition sources in the reference spatial unit are determined to have a target candidate intensity greater than the target intensity threshold, a risk state quantity greater than the risk threshold, or a view complementarity factor greater than the complementarity threshold, and the number of source-level retained points of the effective point cloud acquisition sources is increased so that the sum of the source-level retained points of each point cloud acquisition source in the reference spatial unit reaches the preset minimum number of retained points of the spatial unit or there are no effective point clouds available for retention; After applying the lower limit constraint on coverage, apply the upper limit constraint on budget: if the sum of the source-level reserved points of each point cloud acquisition source in the reference spatial unit is greater than the preset maximum reserved point number of the spatial unit, reduce the source-level reserved point number of each point cloud acquisition source according to the source allocation weight or point-level priority, and determine the sum of the corrected source-level reserved points as the final spatial unit-level reserved point number of the reference spatial unit; the final spatial unit-level reserved point number is less than or equal to the preset maximum reserved point number of the spatial unit. After applying the budget cap constraint, the overall budget constraint is applied: if the sum of the final spatial unit-level reserved points for all spatial units exceeds the global budget points, a corresponding reduction priority is calculated based on the risk state quantity, source average reliability, and coverage overlap of each specific spatial unit. The spatial unit-level reserved points for each specific spatial unit are then reduced based on these reduction priorities until the sum of the final spatial unit-level reserved points for all spatial units satisfies the global budget points. Here, a specific spatial unit is defined as a spatial unit whose risk state quantity is less than a risk state threshold, whose source average reliability is less than a spatial unit reliability threshold, and whose coverage overlap is greater than a redundancy threshold.

[0067] In some embodiments, a lower limit threshold for source reliability is set. and source normal allocation threshold And 0≤ < ≤1. During the execution of the reliability lower limit constraint, when the source reliability assessment value of the point cloud acquisition source is less than 1. When the source reliability assessment value of the point cloud acquisition source is greater than or equal to zero, the source-level reserved point count of the point cloud acquisition source is set to zero; when the source reliability assessment value of the point cloud acquisition source is greater than or equal to zero, the source-level reserved point count of the point cloud acquisition source is set to zero. and less than Furthermore, the complementary factor of the viewpoint is greater than the complementary threshold. When the source reliability assessment value of the point cloud acquisition source is greater than or equal to the preset minimum source-level reserved point number, the source-level reserved point number is set to the minimum. At this time, there is no need to correct the number of source-level reserved points for the point cloud acquisition source; that is, the point cloud acquisition source participates in the normal source quota allocation. Among these, , and Instead of setting fixed values ​​across scenes, the source reliability assessment value and view complementarity factor are normalized to [0, 1] and determined by the center parameter and scale parameter of the corresponding Gaussian membership function and the consequent parameter of the TSK fuzzy inference system. The Gaussian membership function parameter is initialized based on key statistics, and the consequent parameter is adjusted by backpropagation of the downstream 3D target detection loss.

[0068] In some embodiments, during the execution of source proportion constraints, if the number of first source-level reserved points is greater than the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit... The product of these factors will correct the number of first-source-level reserved points to the number of second-source-level reserved points, which is the reference space unit. Number of spatial unit level reserved points With respect to the preset source ratio limit product As a second source of reserved points. The number of valid point cloud acquisition sources within the reference spatial unit, the minimum source quota required for other valid sources, and the verification results under a fixed total point count budget are determined; when there are within the reference spatial unit There must be one valid source, and each of the other sources must account for at least one unit of the space unit budget. hour, Not greater than . and The number of valid point cloud acquisition sources within a spatial unit, the number of valid points from each source, and the number of Top-k retained points within a voxel are dynamically determined without setting a uniform constant across spatial units.

[0069] In some embodiments, during the execution of the coverage lower limit constraint, if the sum of the source-level retained points of each point cloud acquisition source in the reference spatial unit is less than the preset minimum retained point number of the spatial unit, the source-level retained point number of effective point cloud acquisition sources with target candidate intensity greater than the target intensity threshold, risk state quantity greater than the risk threshold, or viewpoint complementarity factor greater than the complementarity threshold is preferentially increased, until the sum of the source-level retained points reaches the preset minimum retained point number or there are no effective point clouds available for retention. Here, having a complementary viewpoint means that the differences in observation direction, visible area, or occlusion relationship of different point cloud acquisition sources for the same spatial unit meet preset complementarity conditions, rather than being determined solely based on the location of the spatial unit.

[0070] In some embodiments, during the execution of the budget ceiling constraint, if the sum of the source-level reserved points of each point cloud acquisition source in the reference spatial unit is greater than the preset maximum reserved point number of the spatial unit, the source-level reserved point number is reduced according to the source allocation weight or point-level priority, so that the corrected source-level reserved point number of each point cloud acquisition source in the reference spatial unit is not greater than the preset maximum reserved point number, and the corrected source-level reserved point number is determined as the final spatial unit-level reserved point number of the reference spatial unit.

[0071] In some embodiments, during the execution of the total budget constraint, if the sum of the final spatial unit-level reserved points of each spatial unit exceeds the global budget points... Calculate excess points For a specific space cell, the risk state quantity is less than the risk state threshold, the average source reliability is less than the space cell reliability threshold, and the coverage overlap is greater than the redundancy threshold. (Risk state threshold, spatial unit reliability threshold, and redundancy are represented by the Gaussian membership function corresponding to the verification data), based on the risk state quantity. Average source reliability assessment value (for a specific spatial unit) (Obtained by averaging the source reliability assessment values ​​of all valid point cloud acquisition sources) and coverage overlap. Calculate reduction priority : (5); in, To ensure that the normalized sum of non-negative weights equals 1, the weights are adjusted through backpropagation using the downstream 3D object detection loss on the validation dataset. This minimizes the impact of the reduction strategy on object detection accuracy while satisfying the global budget constraint. The design principle of this formula is that spatial cells with lower risk, lower source reliability, and higher coverage overlap have higher reduction priority and should be prioritized for reduction when the global budget is exceeded.

[0072] After calculating the reduction priority of a specific spatial unit, in each round, the specific spatial unit with the highest current reduction priority is selected, and its spatial unit-level reserved points are reduced by a preset number of points or a preset percentage. Then, the spatial unit-level reserved points and excess points of each spatial unit are recalculated. until excess points Less than or equal to 0. If the number of reserved points at the spatial unit level of all spatial units has been reduced to the preset minimum number of reserved points, but the sum of the final number of reserved points at the spatial unit level of all spatial units is still greater than the global budget number. If the emergency compression mode is activated, all point cloud points are sorted globally from lowest to highest point priority, and then discarded one by one starting from the lowest priority, until the total number of remaining points does not exceed the global budget. .

[0073] Understandably, by using coverage lower limit constraints and budget upper limit constraints, we can prevent effective spatial units from being compressed into empty spaces, while preventing high-density areas from occupying too many global budget points. By using source proportion constraints and coverage lower limit constraints, we can maintain the balance of source coverage in multi-source collaborative sensing and reduce the problem of single source dominating compression results.

[0074] In some embodiments, determining the point cloud data to be retained based on the corrected source-level retained points and the point-level priority ranking results of the point cloud data collected by each point cloud acquisition source includes: selecting candidate point cloud data from the reference spatial unit with a number not exceeding the corrected source-level retained points of the reference point cloud acquisition source based on the point-level priority ranking results of the point cloud data collected by the reference point cloud acquisition source; and using the candidate point cloud data as the point cloud data to be retained corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0075] In some embodiments, for point cloud data collected by a reference point cloud acquisition source, the top N (N less than or equal to the corrected source-level retained points) point cloud data can be selected as candidate point cloud data from the point-level priority sorting results of the point cloud data collected by the reference point cloud acquisition source, based on the corrected source-level retained points of the reference point cloud acquisition source. The candidate point cloud data are then used as the point cloud data to be retained corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0076] The multi-source point cloud spatial cell-level dynamic compression method for unmanned systems provided in this application utilizes a processing device that can perform source reliability assessment, budget point allocation, and selection of point clouds to be retained at a local or edge node, and then send the compressed point cloud to the receiving end or fusion end (cooperative sensing processing node). The compressed point cloud is sent to the cooperative sensing processing node via a communication link, which includes at least one of vehicle-to-vehicle communication, vehicle-to-infrastructure communication, robot-to-robot communication, edge computing communication, or wireless local area communication.

[0077] In some embodiments, compressed point cloud data can be input into a 3D target detection network to output the 3D position, size, and category of vehicles, pedestrians, obstacles, or other targets; compressed point cloud data can also be input into a point cloud semantic segmentation network to output the semantic category of each point cloud point or spatial unit; input into a target tracking network to obtain target identification, motion trajectory, and motion state; compressed point cloud data can also be input into an occupancy mapping network to generate an occupancy grid or occupancy probability representing passable and obstacle areas; or compressed point cloud data can be input into a multi-source fusion network for joint sensing with images, millimeter-wave radar, or other point cloud data. The specific structure of the downstream network is not limited to the spatial unit-level and source-level budget allocation mechanisms of this application.

[0078] In this embodiment, the three-dimensional space formed by point cloud data collected from multiple point cloud acquisition sources is divided into multiple spatial units. Based on the information of the point cloud data and the source metadata of the point cloud acquisition sources, the spatial unit state parameters corresponding to each spatial unit are calculated. Based on the spatial unit state parameters of the reference spatial unit and the global budget points, the number of retained points at the spatial unit level is determined. For each point cloud acquisition source corresponding to the point cloud data in each spatial unit, a source reliability assessment value is calculated based on the corresponding source metadata. Based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit, the number of retained points at the source level is calculated. Based on the number of retained points at the source level and the point-level priority ranking result of the point cloud data collected by each point cloud acquisition source, the point cloud data to be retained is determined and merged to obtain compressed point cloud data. This method enables the point cloud compression process to not only focus on point-level priority but also to perform retention point budget control at the spatial unit level and the point cloud acquisition source level, thereby improving spatial coverage balance and source reliability under the condition of limited global budget points.

[0079] To verify the technical effectiveness of the multi-source point cloud spatial cell-level dynamic compression method for unmanned systems provided in this application, multi-source single-time experiments were conducted on the OPV2V dataset. OPV2V includes over 70 scenes, 11,464 frames of point cloud data, and 232,913 3D vehicle bounding boxes; time synchronization and coordinate registration were completed between the point cloud acquisition sources of each vehicle. Simultaneously, a single-device continuous-time point cloud experiment was conducted on the NuScenes dataset, selecting a keyframe and the six frames before and after it as input. The experiments were implemented using PyTorch and CenterPoint. The OPV2V voxel size was [0.4 m, 0.4 m, 0.2 m], and the NuScenes voxel size was [0.075 m, 0.075 m, 0.2 m]. Each voxel retained a maximum of 32 points. The Adam optimizer and a single-cycle learning rate strategy were used, and training was performed using four NVIDIA V100 GPUs.

[0080] In the point compression comparison experiment, the dropout ratio was set to 0.5, i.e., 50% of the original point cloud points were retained; in the limited traffic experiment, the traffic ratio was set to 0.0002 to 0.2. Unified random dropout, PointNet, and the feature learning network FLN were used as preprocessing baselines, and PointPillars and CenterPoint were used as downstream 3D object detection networks, respectively.

[0081] Evaluation metrics include preprocessing latency and the average accuracy (AP@0.5) and AP@0.7 for 3D target detection at intersection-union (IU) thresholds of 0.5 and 0.7, respectively. Experimental results for each method on the OPV2V default scene and the Culver scene are shown in Table 1.

[0082] Table 1. Experimental results of different methods on the OPV2V default scenario and the Culver scenario. ; As shown in Table 1, under the CenterPoint detection network, a dropout ratio of 0.5, and the default OPV2V scenario, the 3D target detection accuracy of the proposed method (0.915) reaches the optimal level in terms of AP@0.5, which is better than FLN (0.901) and PointNet (0.898), and significantly better than unified random dropout (0.791). In terms of AP@0.7, the proposed method (0.843) is on par with PointNet (0.843) and slightly lower than FLN (0.850). In the Culver scenario, the AP@0.5 and AP@0.7 of the proposed method are 0.812 and 0.697, respectively, which are slightly lower than FLN (AP@0.5 is 0.815, AP@0.7 is 0.707) and PointNet (AP@0.5 is 0.819, AP@0.7 is 0.695), but the differences are all within 0.007, which is within a small fluctuation range. The preprocessing latency of the method in this application is only 6ms, far lower than FLN's 34ms and PointNet's 29ms, and only slightly higher than the uniform random dropout of 2ms. Overall, this method reduces the preprocessing latency by approximately 80% while maintaining detection accuracy comparable to or even better than existing state-of-the-art methods (FLN, PointNet), achieving a balance between detection accuracy and efficiency. Furthermore, when using PointPillars as the downstream detection network and a 0.5 dropout ratio, this method maintains high accuracy at AP@0.5 (0.914 and 0.809 in the two scenarios) and AP@0.7 (0.827 and 0.679 in the two scenarios) in both the default OPV2V scenario and the Culver scenario, indicating that this method has good versatility and adaptability to different downstream detection networks.

[0083] Further ablation experiments were conducted on the point selection module and distance-aware convolution, and the results were compared when the dropout ratio γ was 0.5 and 0. The ablation experiments used PointPillars as the base detector, and the experimental results are shown in Table 2.

[0084] Table 2 Ablation Experiment Results ; As shown in Table 2, in the default OPV2V scenario, the AP@0.5 and AP@0.7 of the basic detector PointPillars are 0.802 and 0.684, respectively. After introducing the point selection module alone, AP@0.5 and AP@0.7 are improved to 0.902 and 0.795, respectively, with improvements of 10.0 and 11.1 percentage points, indicating that the point selection module significantly improves detection accuracy by effectively filtering key point cloud points. After introducing distance-aware convolution alone, AP@0.5 and AP@0.7 are improved to 0.872 and 0.781, respectively, verifying the gain effect of distance information on detection performance. When the two are used together (the complete method of this application), AP@0.5 and AP@0.7 are further improved to 0.914 and 0.827, respectively, both better than the configuration of any single module, indicating that there is a good synergistic gain effect between the point selection module and the distance-aware convolution. In the Culver scenario, the AP@0.5 and AP@0.7 of the complete method of this application are 0.809 and 0.679, respectively, when the discard ratio γ=0.5. When γ is reduced to 0.0 (i.e. no point cloud is discarded), the accuracy is improved to 0.832 and 0.729, respectively. This verifies the impact of compression degree on detection accuracy. That is, a higher retention rate (lower discard ratio) is beneficial to retain more effective information. However, even at a 50% discard ratio, the method of this application can still maintain high detection accuracy, which demonstrates its ability to effectively maintain perception performance while significantly compressing point cloud data.

[0085] This application provides a multi-source point cloud spatial unit-level dynamic compression device for unmanned systems, such as... Figure 2 As shown, the multi-source point cloud spatial unit-level dynamic compression device 200 for unmanned systems includes: The spatial partitioning module 201 is used to divide the three-dimensional space composed of point cloud data collected by multiple point cloud acquisition sources into multiple spatial units. The spatial state parameter calculation module 202 is used to calculate the spatial unit state parameters corresponding to each spatial unit based on the information of the point cloud data and the source metadata of the point cloud acquisition source; the spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. The spatial unit budget allocation module 203 is used to determine the number of spatial unit-level retained points of the reference spatial unit based on the spatial unit state parameters and the global budget points of the reference spatial unit; the number of spatial unit-level retained points represents the number of point clouds that the reference spatial unit is allowed to retain after compression; the reference spatial unit can be any one of the spatial units. The source reliability assessment value calculation module 204 is used to calculate the source reliability assessment value for each point cloud acquisition source corresponding to the point cloud data in each spatial unit, based on the corresponding source metadata. The source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. The source-level budget allocation module 205 is used to calculate the number of source-level reserved points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of spatial unit-level reserved points of the reference spatial unit; the number of source-level reserved points represents the number of point clouds acquired by the reference point cloud acquisition source that is allowed to be retained in the reference spatial unit; the reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit. The point cloud selection and compression module 206 is used to determine the point cloud data to be retained based on the number of points to be retained at the source level and the point priority sorting results of the point cloud data collected by each point cloud acquisition source, and to merge them to obtain compressed point cloud data.

[0086] In some embodiments, the space unit budget allocation module 203 is further configured to: calculate the retention weight of the reference space unit based on the space unit state parameters of the reference space unit; calculate the ratio of the retention weight of the reference space unit to the sum of the retention weights of each space unit, and multiply it by the global budget points to obtain the initial space unit-level retention points of the reference space unit; and truncate the initial space unit-level retention points using the preset minimum retention points and preset maximum retention points of the space unit to obtain the space unit-level retention points of the reference space unit.

[0087] In some embodiments, the spatial cell state parameters further include point density, target candidate strength, geometric abruptness, and risk state quantity; the spatial cell budget allocation module 203 is further configured to: normalize the number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity of the reference spatial cell respectively, to obtain normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity; and perform weighted summation of the normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity to obtain the retention weight of the reference spatial cell.

[0088] In some embodiments, the multi-source point cloud spatial cell-level dynamic compression device 200 for unmanned systems further includes a weight adjustment unit, which is used to increase the retention weight of the reference spatial cell when the target candidate intensity of the reference spatial cell is greater than the target intensity threshold, or the geometric abruptness of the reference spatial cell is greater than the geometric abruptness threshold; and to decrease the retention weight of the reference spatial cell when the reference spatial cell contains only a dense background and the coverage overlap is greater than the coverage overlap threshold.

[0089] In some embodiments, the source-level budget allocation module 205 is further configured to calculate the source allocation weight of the reference point cloud acquisition source in the reference spatial unit based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit, the view complementarity factor, and the effective state value of the point cloud data acquired by the reference point cloud acquisition source; the view complementarity factor is generated based on the observation direction difference, incident angle difference, or occlusion relationship of the point cloud acquisition sources forming complementary viewpoints in the reference spatial unit; the source allocation weight is calculated as the ratio of the source allocation weight to the sum of the source allocation weights of each point cloud acquisition source in the reference spatial unit, and multiplied by the number of spatial unit-level reserved points in the reference spatial unit to obtain the number of source-level reserved points corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0090] In some embodiments, the point cloud selection and compression module 206 is further configured to: correct the number of source-level retained points and the number of spatial unit-level retained points based on budget allocation constraints to obtain the corrected number of source-level retained points; during the correction process, the sum of the number of source-level retained points does not exceed the number of spatial unit-level retained points; and determine the point cloud data to be retained based on the corrected number of source-level retained points and the point-level priority sorting result of the point cloud data collected by each point cloud acquisition source.

[0091] In some embodiments, the budget allocation constraints include, sequentially executed, a reliability lower limit constraint, a source proportion constraint, a coverage lower limit constraint, a budget upper limit constraint, and a total budget constraint; the point cloud selection and compression module 206 is further configured to: First, a reliability lower limit constraint is applied: if the source reliability assessment value is less than the reliability lower limit threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is set to zero; if the source reliability assessment value is greater than or equal to the reliability lower limit threshold and less than the source normal allocation threshold, and the view complementarity factor of the reference point cloud acquisition source in the reference spatial unit is greater than the complementarity threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is corrected to the first source-level reserved point number; the first source-level reserved point number is the preset minimum source-level reserved point number. After applying the reliability lower limit constraint, the source proportion constraint is applied: if the number of first source-level reserved points is greater than the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit, the number of first source-level reserved points is adjusted to the number of second source-level reserved points; the number of second source-level reserved points is the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit. After implementing the source proportion constraint, the coverage lower limit constraint is implemented: the effective point cloud acquisition sources in the reference spatial unit are determined to have a target candidate intensity greater than the target intensity threshold, a risk state quantity greater than the risk threshold, or a view complementarity factor greater than the complementarity threshold, and the number of source-level retained points of the effective point cloud acquisition sources is increased so that the sum of the source-level retained points of each point cloud acquisition source in the reference spatial unit reaches the preset minimum number of retained points of the spatial unit or there are no effective point clouds available for retention; After applying the lower limit constraint on coverage, apply the upper limit constraint on budget: if the sum of the source-level reserved points of each point cloud acquisition source in the reference spatial unit is greater than the preset maximum reserved point number of the spatial unit, reduce the source-level reserved point number of each point cloud acquisition source according to the source allocation weight or point-level priority, and determine the sum of the corrected source-level reserved points as the final spatial unit-level reserved point number of the reference spatial unit; the final spatial unit-level reserved point number is less than or equal to the preset maximum reserved point number of the spatial unit. After the budget cap constraint is executed, the overall budget constraint is executed: if the sum of the final spatial unit-level reserved points of each spatial unit exceeds the global budget points, the corresponding reduction priority is calculated based on the risk state quantity, source average reliability, and coverage overlap of each specific spatial unit, and the spatial unit-level reserved points of the specific spatial unit are reduced based on each reduction priority until the sum of the final spatial unit-level reserved points of each spatial unit satisfies the global budget points; whereby, a specific spatial unit is a spatial unit whose risk state quantity is less than the risk state threshold, whose source average reliability is less than the spatial unit reliability threshold, and whose coverage overlap is greater than the redundancy threshold.

[0092] In some embodiments, the point cloud selection and compression module 206 is further configured to: select candidate point cloud data from the reference spatial unit based on the point-level priority sorting result of the point cloud data collected by the reference point cloud acquisition source, with a number not exceeding the number of source-level retained points corrected by the reference point cloud acquisition source; and use the candidate point cloud data as the point cloud data to be retained corresponding to the reference point cloud acquisition source in the reference spatial unit.

[0093] The description of the above device embodiments is similar to the description of the above method embodiments, and has the same or similar beneficial effects as the method embodiments. For technical details not disclosed in the device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0094] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of this application can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of this application.

[0095] The methods described in the embodiments of this application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code downloaded over a network that is originally stored in a remote recording medium or a non-transitory machine-readable medium and will be stored in a local recording medium. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0096] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0097] The above embodiments are only used to illustrate the embodiments of this application, and are not intended to limit the embodiments of this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of this application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of this application, and the patent protection scope of the embodiments of this application should be defined by the claims.

Claims

1. A method for dynamic compression of multi-source point cloud spatial units for unmanned systems, characterized in that, include: The three-dimensional space composed of point cloud data collected from multiple point cloud acquisition sources is divided into multiple spatial units; The point cloud data acquisition source is deployed on an unmanned system; Based on the information of the point cloud data and the source metadata of the point cloud acquisition source, calculate the spatial unit state parameters corresponding to each spatial unit. Spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. The number of points to be retained at the spatial unit level for the reference spatial unit is determined based on the spatial unit state parameters and the global budget points. The number of points to be retained at the spatial unit level represents the number of point clouds that the reference spatial unit is allowed to retain after compression. The reference spatial unit can be any one of the spatial units. For each point cloud acquisition source corresponding to the point cloud data in each spatial unit, a source reliability assessment value is calculated based on the corresponding source metadata; the source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. The number of source-level retained points is calculated based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit. The number of points retained at the source level represents the number of point clouds collected by the reference point cloud acquisition sources that are allowed to be retained in the reference spatial unit; the reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit; Based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, the point cloud data to be retained is determined and merged to obtain compressed point cloud data.

2. The method according to claim 1, characterized in that, The process of determining the number of reserved points at the spatial cell level for the reference spatial cell based on the spatial cell state parameters and the global budget points includes: The retention weight of the reference space element is calculated based on the space element state parameters of the reference space element. Calculate the ratio of the retention weight of the reference spatial cell to the sum of the retention weights of all spatial cells, and multiply it by the global budget points to obtain the initial spatial cell-level retention points of the reference spatial cell; By using the preset minimum and maximum number of reserved points of the spatial unit, the number of reserved points at the initial spatial unit level is truncated to obtain the number of reserved points at the spatial unit level of the reference spatial unit.

3. The method according to claim 2, characterized in that, The spatial unit state parameters also include point density, target candidate strength, geometric abruptness, and risk state quantity; The calculation of the retention weight of the reference space cell based on the space cell state parameters of the reference space cell includes: The number of point cloud acquisition sources, coverage overlap, point density, target candidate intensity, geometric abruptness, and risk state quantity of the reference spatial unit are normalized respectively to obtain the normalized number of point cloud acquisition sources, coverage overlap, point density, target candidate intensity, geometric abruptness, and risk state quantity. The retention weights of the reference spatial units are obtained by weighting and summing the number of normalized point cloud acquisition sources, coverage overlap, point density, target candidate strength, geometric abruptness, and risk state quantity.

4. The method according to claim 3, characterized in that, Also includes: If the target candidate strength of the reference space cell is greater than the target strength threshold, or the geometric abruptness of the reference space cell is greater than the geometric abruptness threshold, increase the retention weight of the reference space cell. When the reference spatial cell contains only a dense background and the coverage overlap is greater than the coverage overlap threshold, the retention weight of the reference spatial cell is reduced.

5. The method according to claim 1, characterized in that, The step of calculating the number of source-level retained points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of retained points at the spatial unit level of the reference spatial unit includes: Based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit, the view complementarity factor, and the effective state value of the point cloud data acquired by the reference point cloud acquisition source, the source allocation weight of the reference point cloud acquisition source in the reference spatial unit is calculated; the view complementarity factor is determined based on the differences in observation direction, incident angle, visible area, or occlusion relationship of different reference point cloud acquisition sources to the same reference spatial unit. Calculate the ratio of the source allocation weight to the sum of the source allocation weights of each point cloud acquisition source in the reference spatial unit, and multiply it by the number of spatial unit-level reserved points in the reference spatial unit to obtain the number of source-level reserved points corresponding to the reference point cloud acquisition source in the reference spatial unit.

6. The method according to claim 1, characterized in that, The step of determining the point cloud data to be retained based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source includes: The number of source-level and spatial unit-level reserved points is adjusted based on budget allocation constraints to obtain the adjusted number of source-level reserved points; during the adjustment process, the sum of the number of source-level reserved points does not exceed the number of spatial unit-level reserved points. Based on the corrected source-level retention points and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, the point cloud data to be retained is determined.

7. The method according to claim 6, characterized in that, The budget allocation constraints include, in sequence, the reliability lower limit constraint, the source proportion constraint, the coverage lower limit constraint, the budget upper limit constraint, and the total budget constraint; The correction of the number of source-level and spatial unit-level reserved points based on budget allocation constraints includes: First, a reliability lower limit constraint is applied: if the source reliability assessment value is less than the reliability lower limit threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is set to zero; if the source reliability assessment value is greater than or equal to the reliability lower limit threshold and less than the source normal allocation threshold, and the view complementarity factor of the reference point cloud acquisition source in the reference spatial unit is greater than the complementarity threshold, the number of source-level reserved points of the reference point cloud acquisition source in the reference spatial unit is corrected to the first source-level reserved point number; the first source-level reserved point number is the preset minimum source-level reserved point number. After applying the reliability lower limit constraint, the source proportion constraint is applied: if the number of first source-level reserved points is greater than the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit, the number of first source-level reserved points is adjusted to the number of second source-level reserved points; the number of second source-level reserved points is the product of the number of spatial unit-level reserved points of the reference spatial unit and the preset source proportion upper limit. After implementing the source proportion constraint, the coverage lower limit constraint is implemented: the effective point cloud acquisition sources in the reference spatial unit are determined to have a target candidate intensity greater than the target intensity threshold, a risk state quantity greater than the risk threshold, or a view complementarity factor greater than the complementarity threshold, and the number of source-level retained points of the effective point cloud acquisition sources is increased so that the sum of the source-level retained points of each point cloud acquisition source in the reference spatial unit reaches the preset minimum number of retained points of the spatial unit or there are no effective point clouds available for retention; After applying the lower limit constraint on coverage, apply the upper limit constraint on budget: if the sum of the source-level reserved points of each point cloud acquisition source in the reference spatial unit is greater than the preset maximum reserved point number of the spatial unit, reduce the source-level reserved point number of each point cloud acquisition source according to the source allocation weight or point-level priority, and determine the sum of the corrected source-level reserved points as the final spatial unit-level reserved point number of the reference spatial unit; the final spatial unit-level reserved point number is less than or equal to the preset maximum reserved point number of the spatial unit. After the budget cap constraint is executed, the overall budget constraint is executed: if the sum of the final spatial unit-level reserved points of each spatial unit exceeds the global budget points, the corresponding reduction priority is calculated based on the risk state quantity, source average reliability, and coverage overlap of each specific spatial unit, and the spatial unit-level reserved points of the specific spatial unit are reduced based on each reduction priority until the sum of the final spatial unit-level reserved points of each spatial unit satisfies the global budget points; whereby, a specific spatial unit is a spatial unit whose risk state quantity is less than the risk state threshold, whose source average reliability is less than the spatial unit reliability threshold, and whose coverage overlap is greater than the redundancy threshold.

8. The method according to claim 6, characterized in that, The step of determining the point cloud data to be retained based on the corrected source-level retained point count and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source includes: Based on the point-level priority sorting results of the point cloud data collected by the reference point cloud acquisition source, candidate point cloud data with a number not exceeding the number of source-level retained points after correction of the reference point cloud acquisition source are selected from the reference spatial unit. The candidate point cloud data is used as the point cloud data to be retained corresponding to the reference point cloud acquisition source in the reference spatial unit.

9. The method according to any one of claims 1, 6, 7, and 8, characterized in that, The point-level priority ranking result is determined based on at least one of the following: distance from point cloud data to the corresponding sensor of point cloud acquisition source, reflection intensity, local point density, local curvature, and target candidate intensity.

10. A multi-source point cloud spatial unit-level dynamic compression device for unmanned systems, characterized in that, include: The spatial partitioning module is used to divide the three-dimensional space composed of point cloud data collected from multiple point cloud acquisition sources into multiple spatial units. The spatial state parameter calculation module is used to calculate the spatial unit state parameters corresponding to each spatial unit based on the information of the point cloud data and the source metadata of the point cloud acquisition source. Spatial unit state parameters include at least the number of point cloud acquisition sources and the coverage overlap. The spatial unit budget allocation module is used to determine the number of spatial unit-level retained points of the reference spatial unit based on the spatial unit state parameters and the global budget points. The number of spatial unit-level retained points represents the number of point clouds that the reference spatial unit is allowed to retain after compression. The reference spatial unit can be any one of the spatial units. The source reliability assessment value calculation module is used to calculate the source reliability assessment value for each point cloud acquisition source corresponding to the point cloud data in each spatial unit, based on the corresponding source metadata. The source reliability assessment value is determined at least based on the time freshness of the point cloud data acquisition and the pose confidence of the point cloud acquisition source. The source-level budget allocation module is used to calculate the number of source-level reserved points based on the source reliability assessment value of the reference point cloud acquisition source in the reference spatial unit and the number of spatial unit-level reserved points of the reference spatial unit. The number of points retained at the source level represents the number of point clouds collected by the reference point cloud acquisition sources that are allowed to be retained in the reference spatial unit; the reference point cloud acquisition source is any point cloud acquisition source corresponding to the point cloud data in the reference spatial unit; The point cloud selection and compression module is used to determine the point cloud data to be retained based on the number of points to be retained at the source level and the point-level priority sorting results of the point cloud data collected by each point cloud acquisition source, and to merge them to obtain compressed point cloud data.

Images