Intelligent region division method based on laser radar and related equipment

Abstract

The invention provides an intelligent region division method based on a laser radar and related equipment. The method comprises the following steps: acquiring a target super voxel unit set obtained by performing multi-scale super voxel division on laser radar point cloud data; performing dynamic growth rule expansion on the ground growth seed points of the target super voxel unit set to generate a ground and non-ground region segmentation result; clustering point clouds in different distance ranges in the non-ground super voxel unit set of the region segmentation result according to an adaptive clustering threshold to generate an obstacle clustering target set; and according to the obstacle clustering target set, boundary fitting is carried out on the obstacle target, and the area boundary of the passable area and the obstacle area is determined. Through the implementation of the scheme of the invention, the problem that the recognition of the obstacle target is unstable in an environment with large density change in the prior art is effectively solved, and the accuracy of intelligent region division is improved.

Description

Technical Field

[0001] The present application relates to the field of laser radar technology, and in particular to a laser radar-based intelligent area division method and related equipment. Background Art

[0002] In areas such as autonomous driving, smart transportation, and intelligent security, environmental perception is a crucial foundation for decision-making and control. LiDAR, due to its high precision, strong anti-interference capabilities, and ability to acquire three-dimensional spatial information, has become a crucial technology for intelligent area segmentation. In autonomous driving scenarios, vehicles use LiDAR point cloud data to identify roads, obstacles, and traversable areas for path planning and obstacle avoidance. For example, on urban roads, vehicles must not only detect static obstacles ahead (such as guardrails and parked vehicles), but also identify dynamic targets (such as pedestrians and cyclists) and complex terrain (such as slopes and potholes) to ensure safe driving. However, existing area segmentation methods have limitations when dealing with complex, multi-scale environments. In particular, when point cloud data density is uneven and terrain is complex, the accuracy of distinguishing between the ground and obstacles decreases, resulting in inaccurate determination of traversable areas and affecting vehicle navigation decisions. Summary of the Invention

[0003] This application provides a lidar-based intelligent area division method and related equipment to solve the problem that related technologies have limitations when facing multi-scale complex environments.

[0004] In a first aspect, the present application provides a laser radar-based intelligent area division method, the laser radar-based intelligent area division method comprising: Obtaining a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data; Generating a ground and non-ground area segmentation result by dynamically expanding the ground growth seed points of the target supervoxel unit set; Clustering point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to an adaptive clustering threshold to generate an obstacle clustering target set; Boundary fitting is performed on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

[0005] Optionally, in a first implementation of the first aspect of the present application, the step of obtaining a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data includes: determining a supervoxel division scale according to the detection performance of the laser radar; Performing regional segmentation on the point cloud data according to the supervoxel division scale to obtain initial supervoxel sets at different scales; constructing a multidimensional feature space according to the spatial coordinates, point cloud normal direction, local curvature and reflection intensity of the initial supervoxel set; generating a supervoxel unit set with geometric feature constraints according to feature differences between adjacent supervoxels in the multidimensional feature space; Obtaining an optimized supervoxel unit set by performing energy function optimization calculation on the supervoxel unit set; Statistical analysis is performed on the point cloud distribution of the optimized supervoxel unit set, and supervoxels that meet the uniform distribution characteristics are screened according to a preset threshold to generate a target supervoxel unit set.

[0006] Optionally, in a second implementation of the first aspect of the present application, the step of generating a ground and non-ground area segmentation result by dynamically expanding the ground growth seed points of the target supervoxel unit set includes: Based on the normal vector distribution and elevation value of the target supervoxel unit set, initial ground seed points are screened by a preset plane fitting residual threshold to generate a set of candidate ground seed points; A dynamic growth rule is constructed based on the normal vector angle and elevation gradient difference between adjacent supervoxel units; Performing region growing iteration on the candidate ground seed point set according to the dynamic growing rule to generate a candidate ground region set; Based on the topological connectivity of the candidate ground area set, isolated areas are eliminated by setting a maximum discontinuous span threshold to generate an optimized ground area segmentation result; The non-ground supervoxel units are reversely labeled according to the ground area segmentation result to generate ground and non-ground area segmentation results.

[0007] Optionally, in a third implementation of the first aspect of the present application, the step of clustering point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to the adaptive clustering threshold to generate an obstacle cluster target set includes: Compensate and correct the point cloud density within different distance ranges based on the distribution characteristics of the lidar scan line to generate a density-corrected non-ground point cloud set; Constructing a dynamic clustering threshold function according to local point cloud density and signal-to-noise ratio parameters in the non-ground point cloud set; Performing multi-scale spatial clustering on the non-ground point cloud set by using the dynamic clustering threshold function to generate an initial obstacle cluster set; Based on the overlapping area volume ratio and density confidence parameters of the initial obstacle cluster set, low-confidence clusters are eliminated through competitive screening rules to generate an obstacle cluster target set.

[0008] Optionally, in a fourth implementation of the first aspect of the present application, the step of performing boundary fitting on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area includes: spatially sorting the point cloud of the obstacle cluster target set according to the topological continuity of the laser radar scan line to generate a scan line aligned obstacle point cloud sequence; According to the geometric feature distribution of the obstacle point cloud sequence, regular obstacles and irregular obstacles are fitted respectively by a multimodal boundary generation algorithm to determine an initial boundary set; Performing type annotation on the initial boundary set according to preset semantic annotation rules; Based on the spatial distribution characteristics of the annotation types, the boundary between the passable area and the obstacle area is smoothly optimized through a gridded occupancy probability model to determine the regional boundary between the passable area and the obstacle area.

[0009] Optionally, in a fifth implementation of the first aspect of the present application, the method further includes: generating a dynamic resource allocation weight matrix according to the supervoxel unit distribution density in the region segmentation result; Perform multi-channel iterative constraints on the expansion rate of the ground growth seed points through the dynamic resource allocation weight matrix to generate regional computing resource thresholds; The clustering threads of the non-ground supervoxel unit set are parallelized and scheduled based on the sub-region computing resource threshold, and a point cloud processing queue of the clustering threads is determined.

[0010] Optionally, in a sixth implementation of the first aspect of the present application, the method further includes: generating a multimodal feature fusion weight matrix based on the sub-regional computing resource threshold and the spatiotemporal synchronization data collected by the multi-source sensors; Performing spatiotemporal alignment optimization on the lidar point cloud data and other modal perception data through the multimodal feature fusion weight matrix to generate a cross-sensor calibration point cloud set; Based on the topological consistency of the cross-sensor calibration point cloud set, a global area occupancy map is constructed through a multi-resolution rasterization probability model; The global area occupancy map is incrementally updated through a dynamic Bayesian network to generate an optimized global traversable area map.

[0011] A second aspect of the present application provides a laser radar-based intelligent area division device, the laser radar-based intelligent area division device comprising: An acquisition module is used to obtain a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data; An expansion module, configured to generate a ground and non-ground area segmentation result by performing dynamic growth rule expansion on the ground growth seed points of the target supervoxel unit set; a clustering module, configured to cluster point clouds within different distance ranges in a non-ground supervoxel unit set of the region segmentation result according to an adaptive clustering threshold, to generate an obstacle clustering target set; The fitting module is used to perform boundary fitting on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

[0012] A third aspect of an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the processor is used to execute a computer program stored on the memory. When the processor executes the computer program, it implements each step of the laser radar-based intelligent area division method provided in the first aspect of the embodiment of the present application.

[0013] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the laser radar-based intelligent area division method provided in the first aspect of the embodiment of the present application are implemented.

[0014] In summary, according to the laser radar-based intelligent area division method and related equipment provided by the present application, a target supervoxel unit set obtained by multi-scale supervoxel division of the laser radar point cloud data is obtained; by dynamically expanding the ground growth seed points of the target supervoxel unit set, a ground and non-ground area segmentation result is generated; according to an adaptive clustering threshold, point clouds within different distance ranges in the non-ground supervoxel unit set of the area segmentation result are clustered to generate an obstacle clustering target set; according to the obstacle clustering target set, boundary fitting of obstacle targets is performed to determine the regional boundary between the passable area and the obstacle area. Through the implementation of the present application, the problem of unstable obstacle target recognition in environments with large density changes in the existing technology is effectively solved, and the accuracy of intelligent area division is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a flow chart of a laser radar-based intelligent area division method provided in an embodiment of the present application; Figure 2 A schematic diagram of a program module of a laser radar-based intelligent area division device provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, features, and advantages of the invention of this application more obvious and easy to understand, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of this application.

[0017] In order to solve the problem that the related technology has limitations when facing multi-scale complex environments, the embodiment of the present application provides an intelligent area division method based on laser radar, such as Figure 1 The following is a flow chart of the laser radar-based intelligent area division method provided in this embodiment. The laser radar-based intelligent area division method includes the following steps: Step 110: Obtain a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data.

[0018] Specifically, in this embodiment, the laser radar point cloud data is first divided into multi-scale supervoxels to obtain a target supervoxel unit set. By setting the supervoxel size within different distance ranges, the near-field area and the far-field area can maintain a reasonable division scale when the point cloud density changes. A multidimensional feature space is constructed using spatial coordinates, point cloud normal direction, local curvature and reflection intensity, and on this basis, the feature differences between adjacent supervoxels are calculated to generate a supervoxel unit set with geometric feature constraints. In order to ensure the uniform distribution characteristics of the point cloud data, the supervoxel unit is adjusted through energy function optimization calculation, and its point cloud distribution is statistically analyzed. Supervoxel units with stable features are screened according to a preset threshold, and finally a target supervoxel unit set is formed.

[0019] In an optional implementation of the present embodiment, the step of obtaining a target supervoxel unit set obtained by performing multi-scale supervoxel division on the laser radar point cloud data includes: determining the supervoxel division scale according to the detection performance of the laser radar; performing regional segmentation on the point cloud data according to the supervoxel division scale to obtain initial supervoxel sets at different scales; constructing a multidimensional feature space according to the spatial coordinates, point cloud normal direction, local curvature and reflection intensity of the initial supervoxel set; generating a supervoxel unit set with geometric feature constraints according to the feature differences between adjacent supervoxels in the multidimensional feature space; obtaining an optimized supervoxel unit set by performing energy function optimization calculation on the supervoxel unit set; performing statistical analysis on the point cloud distribution of the optimized supervoxel unit set, and screening supervoxels that meet the uniform distribution characteristics according to a preset threshold to generate a target supervoxel unit set.

[0020] Specifically, in this embodiment, a laser radar (LiDAR) acquires three-dimensional environmental data in an autonomous vehicle scenario. The point cloud data returned after the laser pulse is emitted provides the raw input for region segmentation. First, based on the LiDAR's detection performance, the supervoxel segmentation scale is determined by measuring parameters such as detection range, angular resolution, and echo intensity. A smaller scale is selected for the near-field region and a larger scale is selected for the far-field region to accommodate the characteristic that point cloud density varies with distance. The point cloud data is then segmented to generate initial supervoxel sets at different scales. This set spatially preserves the basic outline information of each target in the environment. Next, the spatial coordinates, point cloud normal direction, local curvature, and reflection intensity of each supervoxel in the initial supervoxel set are extracted. The spatial coordinates represent the point's position in three-dimensional space, the normal direction describes the point cloud surface orientation, the local curvature characterizes the curvature of the point cloud surface, and the reflection intensity reflects the echo characteristics of the laser signal. These parameters constitute a multidimensional feature space, providing a rich feature description for subsequent processing. By comparing the feature differences between adjacent supervoxels, a supervoxel unit set with geometric feature constraints is generated. In order to quantify the feature differences between adjacent supervoxel units, a key energy function is designed, whose mathematical expression is: , in, It represents the geometric consistency cost between supervoxel unit i and supervoxel unit j, and represent the normal vectors of supervoxel units i and j respectively, and represent the local curvature, is the Euclidean distance between the centers of two cells, is the difference in reflection intensity between the two units, , , To adjust the parameters, a formula was obtained through extensive statistical analysis and experiments. This formula is used to measure the differences in geometric and optical properties between adjacent supervoxel units, and the impact of long-distance data errors is reduced through distance normalization and exponential decay. On this basis, an energy function is optimized for the supervoxel unit set, and units with higher energy values ​​are segmented and adjusted to obtain an optimized supervoxel unit set. The point cloud distribution within each unit is statistically analyzed, and the distribution variance is calculated. Supervoxels with uniform distribution and complete edge features are selected based on a preset threshold to generate a target supervoxel unit set. This set can be used in autonomous driving to distinguish roads, obstacles, and other key areas, providing a stable and structured description of the environment for subsequent path planning and obstacle avoidance.

[0021] Step 120 : Generate a ground and non-ground area segmentation result by performing dynamic growth rule expansion on the ground growth seed points of the target supervoxel unit set.

[0022] Specifically, in this embodiment, based on the target supervoxel unit set, supervoxels that meet specific elevation thresholds and normal vector constraints are selected as ground growth seed points, and the expansion range of the ground area is controlled by setting growth rules. Combined with the normal vector distribution, elevation gradient, and topological connectivity, the growth direction is dynamically calculated, and the ground area is gradually expanded along the optimal path. When the height difference between adjacent supervoxels exceeds the set threshold, or the normal vector changes exceed the specified range, the growth of the current area is stopped. Multiple independent ground areas are merged by setting a discontinuous span threshold to eliminate the fracture phenomenon in the ground area, and finally complete the division of the ground and non-ground areas.

[0023] In an optional implementation of this embodiment, the step of generating ground and non-ground area segmentation results by dynamically expanding the ground growth seed points of the target supervoxel unit set includes: based on the normal vector distribution and elevation value of the target supervoxel unit set, screening the initial ground seed points by a preset plane fitting residual threshold to generate a candidate ground seed point set; constructing a dynamic growth rule according to the normal vector angle and elevation gradient difference of adjacent supervoxel units; performing regional growth iteration on the candidate ground seed point set according to the dynamic growth rule to generate a candidate ground area set; based on the topological connectivity of the candidate ground area set, eliminating isolated areas by setting a maximum discontinuous span threshold to generate an optimized ground area segmentation result; and reversely marking the non-ground supervoxel units according to the ground area segmentation result to generate a ground and non-ground area segmentation result.

[0024] Specifically, in this embodiment, based on the normal vector distribution and elevation value information of each unit in the target supervoxel unit set, the point cloud can be plane-fitted using a preset plane fitting residual threshold to screen out initial seed points that meet the ground features. This process uses point cloud data collected by lidar in a vehicle autonomous driving scenario to generate a set of candidate ground seed points. Subsequently, a dynamic growth rule is constructed based on the angle change and elevation gradient difference between the normal vectors of adjacent supervoxel units, and regional growth iterations are performed on the candidate seed point set. This process determines the geometric continuity and elevation consistency between supervoxel units to achieve gradual expansion and connectivity identification of the ground area. In order to quantify the connectivity of regional growth, the dynamic growth judgment function can be expressed as: , in, represents the angle difference between the normal vectors of adjacent supervoxel units (in degrees), is the preset critical angle threshold, represents the elevation gradient difference between adjacent cells (in meters), is the critical elevation difference, r represents the Euclidean distance between the centers of two units, is the maximum allowed distance, 、 and is the shape control parameter, and It is used to determine the connectivity of region growing. When the function value is greater than or equal to the preset threshold, the supervoxel units are considered connected, thereby generating a set of candidate ground regions. Next, the topological connectivity of the candidate ground region set is detected, and isolated or discontinuous regions are eliminated based on the set maximum discontinuous span threshold. This process ensures that the obtained ground region has high coherence and integrity. Finally, using the determined ground region segmentation results, the parts of the original target supervoxel unit set that are not divided into the ground are reversely marked to obtain the final segmentation results of the ground and non-ground areas. In vehicle autonomous driving scenarios, such as on urban roads, accurately separating the road surface and obstacles is of great significance for path planning. This technical solution provides more accurate environmental perception data for the autonomous driving system by deeply analyzing the normal vectors and elevation information between supervoxel units and adopting a method that combines dynamic growth with topological connectivity detection. At the same time, the formula is used to dynamically determine the connectivity of region growing, providing a scientific basis for ground recognition in complex terrain.

[0025] It should be noted that the shape control parameter refers to the exponential parameter used to adjust the nonlinear attenuation or amplification effect, including 、 and Three parameters, which respectively adjust the effects of angle difference, local elevation difference and the distance between two supervoxel units. These parameters are mainly used to control the nonlinear response of different feature differences to the overall connectivity evaluation when calculating the connectivity between adjacent supervoxel units, that is, in the process of judging the region growth, different weights are given to the normal vector difference and the degree of change of the elevation gradient. For example, if the effect of angle difference needs to be more sensitive, then The value of will be large, so that when the difference in normal vectors between supervoxel units exceeds the preset critical angle, the connectivity decays rapidly. Similarly, and They respectively determine the degree to which elevation difference and distance control connectivity. These shape control parameters are not directly measured by lidar sensors, but are obtained through statistical analysis and experimental optimization of a large amount of actual autonomous driving scenario data. Their generation process often relies on experimental debugging and parameter optimization algorithms. For example, in urban road environments, by collecting representative point cloud data and comparing the coherence and accuracy of region growing under different parameter configurations, methods such as grid search, Bayesian optimization, or genetic algorithms can be used to determine a set of parameter values ​​that can effectively distinguish between continuous and discontinuous regions in specific scenarios.

[0026] Step 130 : Cluster the point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to the adaptive clustering threshold to generate an obstacle cluster target set.

[0027] Specifically, in this embodiment, based on the point cloud density distribution characteristics within different distance ranges, the local point cloud density gradient is calculated for a supervoxel set in the non-ground area, and a density distribution model is constructed based on the density change rate. Using a non-uniform distance metric, the distances between adjacent supervoxels are nonlinearly scaled, and a multi-segment density threshold is set, so that a smaller clustering threshold is used in close-range areas, while a larger clustering threshold is used in distant areas to accommodate changes in point cloud density. Based on dynamic connectivity analysis, a search is conducted within the non-ground supervoxel set for groups of supervoxel units that meet connectivity constraints, and the shortest path cost is calculated based on the connectivity weighted graph, ultimately demarcating a set of obstacle cluster targets. Furthermore, adjacent feature inconsistencies are detected for boundary supervoxel units, and cluster boundaries are adjusted based on the normal vector change rate, spatial gradient, and point cloud reflectance characteristics to eliminate possible erroneous clustering.

[0028] In an optional implementation of this embodiment, the step of clustering point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to an adaptive clustering threshold to generate an obstacle cluster target set includes: compensating and correcting the point cloud density within different distance ranges according to the distribution characteristics of the laser radar scan line to generate a density-corrected non-ground point cloud set; constructing a dynamic clustering threshold function based on the local point cloud density and signal-to-noise ratio parameters in the non-ground point cloud set; performing multi-scale spatial clustering on the non-ground point cloud set using the dynamic clustering threshold function to generate an initial obstacle cluster set; and eliminating low-confidence clusters through competitive screening rules based on the overlapping area volume ratio and density confidence parameter of the initial obstacle cluster set to generate an obstacle cluster target set.

[0029] Specifically, in this embodiment, in the vehicle autonomous driving scenario, the point cloud data collected by the lidar exhibits uneven density distribution within different distance ranges. Therefore, it is necessary to compensate and correct the point cloud density in the near field and far field based on the distribution characteristics of the lidar scan line to obtain a density-corrected non-ground point cloud set. In this process, a distance-dependent correction factor is applied to the point cloud density using the lidar's angular resolution and scan line interval. For example, the compensation factor may be greater than 1 in areas less than 30 meters away, while the compensation factor is less than 1 in long-distance areas, thereby compensating for the lack of sparse far-field data. Next, a dynamic clustering threshold function is constructed from the density-corrected non-ground point cloud set based on the statistical characteristics of the local point cloud density and signal-to-noise ratio to determine the connection distance between point clouds in different areas, thereby adapting to the characteristics of near-dense and far-sparse. The dynamic clustering threshold function is used to dynamically calculate the threshold used for clustering in different distance areas. Its purpose is to automatically adjust the critical distance for aggregation between point clouds based on changes in local data density and signal-to-noise ratio. Subsequently, the density-corrected non-ground point cloud set is subjected to multi-scale spatial clustering using the dynamic clustering threshold function to generate an initial set of obstacle clusters. This clustering process groups point clouds that meet the dynamic threshold criteria to form candidate obstacle regions. Furthermore, based on the volume fraction of overlapping regions between clusters within the initial obstacle cluster set and the density confidence parameters within each cluster, a competitive screening rule is employed to remove overlapping clusters with low density confidence, thereby generating a final set of obstacle target clusters. For example, in an urban road environment, when two obstacle clusters partially overlap, redundant or spurious clusters can be removed by calculating the ratio of the overlapping volume to their respective total volumes, combined with local point cloud density assessment, ensuring the accuracy and stability of the final obstacle target set. This entire process, through the dynamic threshold function and competitive screening rule, enables multi-scale and adaptive clustering of point cloud data at varying distances and signal-to-noise ratios, providing a solid foundation for subsequent obstacle boundary fitting and region annotation.

[0030] Step 140: perform boundary fitting on the obstacle targets according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

[0031] Specifically, in this embodiment, a scan line topology is constructed for a set of obstacle cluster targets, and an appropriate boundary fitting method is selected based on the geometric morphology of the obstacle point cloud. For regular obstacle targets, a linear fitting method is used to optimize the point cloud boundary using the least squares method, ensuring that the fitting curve best fits the target contour. For irregular obstacle targets, an Alpha Shape algorithm is used to generate a curved boundary. The fitting curvature is adjusted by setting parameters to accommodate the morphological characteristics of different obstacle types. Based on boundary fitting, the regions are classified and labeled based on reflection intensity information and the length, width, and height dimensions of the obstacle. Ultimately, the boundaries between the traversable and obstructed areas are determined, completing the entire intelligent region segmentation process.

[0032] In an optional implementation of this embodiment, the steps of performing boundary fitting on obstacle targets based on the obstacle cluster target set and determining the regional boundary between the passable area and the obstacle area include: spatially sorting the point cloud of the obstacle cluster target set based on the topological continuity of the lidar scan line to generate a scan-line aligned obstacle point cloud sequence; fitting regular obstacles and irregular obstacles separately based on the geometric feature distribution of the obstacle point cloud sequence using a multimodal boundary generation algorithm to determine an initial boundary set; annotating the initial boundary set by type based on preset semantic annotation rules; and smoothing and optimizing the boundary between the passable area and the obstacle area using a rasterized occupancy probability model based on the spatial distribution characteristics of the annotation types to determine the regional boundary between the passable area and the obstacle area.

[0033] Specifically, in this embodiment, the point clouds in the acquired obstacle cluster target set are spatially sorted based on the topological continuity of the laser radar scan lines. This process is based on the continuity of the scan lines and the relative positional relationship between each point cloud in the scan sequence. The start and end positions of the scan lines and the overlapping parts between adjacent scan lines are analyzed to generate an obstacle point cloud sequence with aligned scan lines. For example, on urban roads in a vehicle autonomous driving scenario, the point cloud data collected by the laser radar will form a series of continuous scan lines along the rotation scan. This continuity ensures that the point clouds of the same obstacle (such as the vehicle ahead or roadside guardrail) are arranged in an orderly manner in the scan sequence. Next, based on the distribution of geometric features in the generated obstacle point cloud sequence, a multimodal boundary generation algorithm is used to fit regular obstacles and irregular obstacles respectively. Regular obstacles can be fitted using a straight line or rectangle, while irregular obstacles are fitted using a curve or contour extraction method. For this purpose, a calculation formula for evaluating the fitting quality of candidate boundary segments is designed, which is expressed as follows: , in, represents the local curvature of the i-th point in the candidate boundary, is the average curvature of the boundary segment, is the standard deviation of the curvature, represents the length deviation between the candidate boundary segment and the standard geometry, is the reference length, n is the number of points in the boundary segment, To regulate parameters and control the impact of curvature deviation, this formula is used to quantify the fitting quality of boundary segments, thereby distinguishing regular from irregular obstacle boundaries. Subsequently, the initially generated boundary set is annotated according to pre-set semantic annotation rules. These semantic annotation rules classify boundaries into traversable and obstruction areas based on characteristics such as boundary geometry, reflection intensity, local curvature, and spatial location. For example, on a major urban road, traversable boundaries typically present a relatively smooth and regular straight line, while obstruction areas may exhibit irregular shapes due to buildings, curbs, or parked vehicles. The annotated boundaries are then smoothed and optimized using a gridded occupancy probability model. This model uses a gridding approach to divide the space into several cells of fixed resolution and updates the occupancy probability within each grid cell. The update rule is based on Bayesian theory and incorporates the distribution information of boundary points within each grid cell to achieve spatial smoothing and enhanced continuity of the boundaries. For example, when a vehicle is traveling on a complex road section, by continuously updating the occupancy probability within each grid cell, the boundary breaks caused by local noise can be eliminated, thereby determining the final boundary between the passable area and the obstacle area. This area boundary serves as an important input for autonomous driving decisions and path planning, ensuring that the vehicle can accurately identify driving channels and potential obstacles in a dynamic environment, thereby providing the system with more stable and reliable environmental perception data.

[0034] In an optional implementation of this embodiment, a dynamic resource allocation weight matrix is ​​generated based on the supervoxel unit distribution density in the regional segmentation result; the expansion rate of the ground growth seed point is multi-channel iteratively constrained by the dynamic resource allocation weight matrix to generate a regional computing resource threshold; based on the regional computing resource threshold, the clustering threads of the non-ground supervoxel unit set are parallelized and scheduled to determine the point cloud processing queue of the clustering threads.

[0035] Specifically, in this embodiment, based on the distribution density information of each supervoxel unit in the regional segmentation result, a dynamic resource allocation weight matrix is ​​constructed by statistically analyzing the local point cloud density of the supervoxel unit. The matrix is ​​used to reflect the density and data uniformity of the supervoxel units in different regions, thereby dynamically adjusting the allocation of computing resources. Specifically, by measuring the density of the point cloud in each supervoxel unit and its distribution variance , and combined with the spatial distance between the supervoxel and the ground growth seed point , generating the following formula: , in, represents the resource allocation weight of supervoxel unit j in region i, is the maximum reference distance set, and To prevent small constants from dividing by zero, this formula is used to generate nonlinear weight adjustments between dense and sparse areas, thereby generating a dynamic resource allocation weight matrix. Next, using the dynamic weight matrix, a multi-channel iterative constraint is imposed on the expansion rate of the ground growth seed points. Each channel determines the computing resource threshold based on the weight information in different regions, and generates a sub-region computing resource threshold. This threshold reflects the processing power requirements of each region in resource scheduling. For example, in an urban autonomous driving scenario, when a vehicle travels in a high-density area, the local supervoxels are dense and the data fluctuations are small. At this time, the corresponding element value in the weight matrix is ​​low, thereby reducing the computing resource threshold of the scoring area to reduce the computational burden of a single processing thread. In low-density areas far from the central road, the corresponding value of the weight matrix is ​​higher, and the sub-region computing resource threshold increases accordingly to ensure the timeliness of data processing. Subsequently, clustering threads for non-ground supervoxel clusters are scheduled in parallel based on regional computing resource thresholds. A scheduling algorithm then divides each clustering thread into multiple processing queues, allocating computing resources to each queue according to a predetermined threshold. Point cloud clustering tasks are then executed in parallel on multi-core or multi-processor platforms, thereby determining the point cloud processing queue for each clustering thread. This process, through the establishment of a dynamic resource allocation weight matrix and multi-channel iterative constraints, ensures efficient and balanced parallel processing within limited computing resources in autonomous vehicle scenarios, regardless of the distribution of point cloud data across different distance ranges, meeting the needs of real-time environmental perception.

[0036] In an optional implementation of this embodiment, a multimodal feature fusion weight matrix is ​​generated based on the regional computing resource threshold and the spatiotemporal synchronization data collected by multi-source sensors; the lidar point cloud data and other modal perception data are spatiotemporally aligned and optimized using the multimodal feature fusion weight matrix to generate a cross-sensor calibration point cloud set; based on the topological consistency of the cross-sensor calibration point cloud set, a global area occupancy map is constructed using a multi-resolution rasterization probability model; the global area occupancy map is incrementally updated using a dynamic Bayesian network to generate an optimized global traversable area map.

[0037] Specifically, in this embodiment, in the vehicle autonomous driving scenario, the lidar point cloud data and the spatiotemporal synchronous data collected by multi-source sensors such as cameras and millimeter-wave radars provide multi-dimensional information for environmental perception. By combining the regional computing resource threshold and the spatiotemporal synchronous data, a set of multimodal feature fusion weight matrices is generated. The core of this matrix is ​​to extract the key features in the data of each sensor, such as the three-dimensional spatial coordinates, reflection intensity and point cloud structure provided by the lidar, the color and texture information provided by the camera, and the target distance and speed information provided by the millimeter-wave radar, and calculate the time difference, spatial position difference and feature descriptor difference of these features respectively. By using the weight matrix, the lidar point cloud data and other modal data are spatiotemporally aligned and optimized, and the matching and calibration algorithm is used to finely align the data collected by different sensors in time and space to generate a cross-sensor calibration point cloud set. This set maintains the integrity of the feature information of each modality while achieving topological consistency between the data, thus laying the foundation for subsequent global modeling. The cross-sensor calibration point cloud collection is then processed using a multi-resolution rasterization probabilistic model. This model maps the continuous 3D point cloud to a fixed-resolution grid. Within each grid, the occupancy state is evaluated through probabilistic calculations to construct a global occupancy map that reflects the spatial distribution characteristics of roads, obstacles, and free areas. Finally, the global occupancy map is incrementally updated using a dynamic Bayesian network that fuses newly acquired data using state transition probabilities. The corresponding Bayesian incremental update formula is expressed as: , in, represents the update probability of the area occupancy state under the k+1th frame data, is the probability of occupying the area at the previous moment, Represents the latest data The conditional probability of the area being occupied is calculated using this formula. This formula is used to incrementally update the global area occupancy map in real time to generate an optimized global traversable area map. This map can reflect the latest information in complex dynamic environments and provide accurate and stable input for autonomous driving decision-making and path planning. For example, in the complex traffic environment during urban rush hour, through the above multimodal data fusion and dynamic updating, the vehicle can obtain the accurate distribution of road traversable areas in real time, thereby ensuring driving safety.

[0038] According to the intelligent region segmentation method based on LiDAR provided by the present application, a target supervoxel unit set is obtained by multi-scale supervoxel segmentation of LiDAR point cloud data; the ground growth seed points of the target supervoxel unit set are dynamically expanded to generate the ground and non-ground region segmentation results; the point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result are clustered according to the adaptive clustering threshold to generate the obstacle cluster target set; the obstacle target is boundary-fitted according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area. Through the implementation of the present application, the problem of unstable obstacle target recognition in environments with large density changes in the existing technology is effectively solved, and the accuracy of intelligent region segmentation is improved.

[0039] Figure 2 The embodiment of the present application provides a laser radar-based intelligent area division device, which can be used to implement the laser radar-based intelligent area division method in the above embodiment. Figure 2 As shown, the intelligent area division device based on laser radar mainly includes: An acquisition module 10 is used to acquire a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data; An expansion module 20 is configured to generate a ground and non-ground area segmentation result by performing dynamic growth rule expansion on the ground growth seed points of the target supervoxel unit set; A clustering module 30 is configured to cluster point clouds within different distance ranges in a non-ground supervoxel unit set of a region segmentation result according to an adaptive clustering threshold, to generate an obstacle clustering target set; The fitting module 40 is used to perform boundary fitting on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

[0040] In an optional implementation of this embodiment, the acquisition module is specifically used to: determine the supervoxel division scale according to the detection performance of the laser radar; perform regional segmentation on the point cloud data according to the supervoxel division scale to obtain initial supervoxel sets at different scales; construct a multidimensional feature space according to the spatial coordinates, point cloud normal direction, local curvature and reflection intensity of the initial supervoxel set; generate a supervoxel unit set with geometric feature constraints according to the feature differences between adjacent supervoxels in the multidimensional feature space; obtain an optimized supervoxel unit set by performing energy function optimization calculation on the supervoxel unit set; perform statistical analysis on the point cloud distribution of the optimized supervoxel unit set, and screen supervoxels that meet the uniform distribution characteristics according to a preset threshold to generate a target supervoxel unit set.

[0041] In an optional implementation of this embodiment, the extension module is specifically used to: based on the normal vector distribution and elevation value of the target supervoxel unit set, screen the initial ground seed points through a preset plane fitting residual threshold to generate a set of candidate ground seed points; construct a dynamic growth rule according to the normal vector angle and elevation gradient difference of adjacent supervoxel units; perform regional growth iteration on the candidate ground seed point set according to the dynamic growth rule to generate a set of candidate ground regions; based on the topological connectivity of the candidate ground region set, eliminate isolated regions by setting a maximum discontinuous span threshold to generate an optimized ground region segmentation result; reversely mark the non-ground supervoxel units according to the ground region segmentation result to generate a ground and non-ground region segmentation result.

[0042] In an optional implementation of this embodiment, the clustering module is specifically used to: compensate and correct the point cloud density within different distance ranges according to the distribution characteristics of the laser radar scan line to generate a density-corrected non-ground point cloud set; construct a dynamic clustering threshold function according to the local point cloud density and signal-to-noise ratio parameters in the non-ground point cloud set; perform multi-scale spatial clustering on the non-ground point cloud set through the dynamic clustering threshold function to generate an initial obstacle cluster set; based on the overlapping area volume ratio and density confidence parameter of the initial obstacle cluster set, eliminate low-confidence clusters through competitive screening rules to generate an obstacle cluster target set.

[0043] In an optional implementation of this embodiment, the fitting module is specifically used to: spatially sort the point cloud of the obstacle cluster target set based on the topological continuity of the lidar scan line to generate a scan-line aligned obstacle point cloud sequence; fit regular obstacles and irregular obstacles separately based on the geometric feature distribution of the obstacle point cloud sequence using a multimodal boundary generation algorithm to determine an initial boundary set; label the initial boundary set by type based on preset semantic labeling rules; and based on the spatial distribution characteristics of the labeling type, smoothly optimize the boundary between the passable area and the obstacle area using a gridded occupancy probability model to determine the regional boundary between the passable area and the obstacle area.

[0044] In an optional implementation of this embodiment, the intelligent region division device further includes a determination module. The determination module is configured to: generate a dynamic resource allocation weight matrix based on the supervoxel unit distribution density in the region segmentation result; perform multi-channel iterative constraints on the expansion rate of ground growth seed points using the dynamic resource allocation weight matrix to generate a regional computing resource threshold; and perform parallel scheduling of clustering threads for a non-ground supervoxel unit set based on the regional computing resource threshold to determine a point cloud processing queue for the clustering threads.

[0045] In an optional implementation of this embodiment, the intelligent region division device further includes: a generation module. The generation module is configured to: generate a multimodal feature fusion weight matrix based on the regional computing resource threshold and the spatiotemporal synchronization data collected by the multi-source sensors; optimize the spatiotemporal alignment of the lidar point cloud data with the other modal perception data using the multimodal feature fusion weight matrix to generate a cross-sensor calibration point cloud set; construct a global region occupancy map using a multi-resolution rasterization probability model based on the topological consistency of the cross-sensor calibration point cloud set; and incrementally update the global region occupancy map using a dynamic Bayesian network to generate an optimized global traversable area map.

[0046] According to the intelligent region segmentation device based on LiDAR provided by the present application, a target supervoxel unit set obtained by multi-scale supervoxel segmentation of LiDAR point cloud data is obtained; the ground growth seed points of the target supervoxel unit set are dynamically expanded to generate a region segmentation result of ground and non-ground areas; point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result are clustered according to an adaptive clustering threshold to generate an obstacle clustering target set; obstacle targets are boundary-fitted according to the obstacle clustering target set to determine the regional boundaries of the passable area and the obstacle area. Through the implementation of the present application, the problem of unstable obstacle target recognition in environments with large density changes in the existing technology is effectively solved, and the accuracy of intelligent region segmentation is improved.

[0047] According to the application plan Figure 3 An electronic device provided in an embodiment of the present application. This electronic device can be used to implement the laser radar-based intelligent area division method in the aforementioned embodiment, mainly including: Memory 301, processor 302, and computer program 303 stored on memory 301 and executable on processor 302. Memory 301 and processor 302 are connected via communication. When processor 302 executes computer program 303, the lidar-based intelligent area division method described in the aforementioned embodiment is implemented. The number of processors can be one or more.

[0048] The memory 301 can be a high-speed random access memory (RAM) memory or a non-volatile memory such as a disk drive. The memory 301 is used to store executable program code. The processor 302 is coupled to the memory 301 .

[0049] Furthermore, the embodiment of the present application also provides a computer-readable storage medium, which can be provided in the electronic device in the above embodiments. The computer-readable storage medium can be the above Figure 3 Memory in the illustrated embodiment.

[0050] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the lidar-based intelligent area segmentation method described in the aforementioned embodiment. Furthermore, the computer-readable storage medium may be a USB flash drive, a mobile hard drive, a read-only memory (ROM), RAM, a magnetic disk, or an optical disk, among other media capable of storing program code.

[0051] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0052] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0053] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A laser radar-based intelligent area division method, characterized in that: include: Obtaining a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data; Generating a ground and non-ground area segmentation result by dynamically expanding the ground growth seed points of the target supervoxel unit set; Clustering point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to an adaptive clustering threshold to generate an obstacle clustering target set; Boundary fitting is performed on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

2. The intelligent area division method based on laser radar according to claim 1, characterized in that: The step of obtaining a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data comprises: determining a supervoxel division scale according to the detection performance of the laser radar; Performing regional segmentation on the point cloud data according to the supervoxel division scale to obtain initial supervoxel sets at different scales; constructing a multidimensional feature space according to the spatial coordinates, point cloud normal direction, local curvature and reflection intensity of the initial supervoxel set; generating a supervoxel unit set with geometric feature constraints according to feature differences between adjacent supervoxels in the multidimensional feature space; Obtaining an optimized supervoxel unit set by performing energy function optimization calculation on the supervoxel unit set; Statistical analysis is performed on the point cloud distribution of the optimized supervoxel unit set, and supervoxels that meet the uniform distribution characteristics are screened according to a preset threshold to generate a target supervoxel unit set.

3. The intelligent area division method based on laser radar according to claim 2, characterized in that: The step of performing dynamic growth rule expansion on the ground growth seed points of the target supervoxel unit set to generate a ground and non-ground area segmentation result includes: Based on the normal vector distribution and elevation value of the target supervoxel unit set, initial ground seed points are screened by a preset plane fitting residual threshold to generate a set of candidate ground seed points; A dynamic growth rule is constructed based on the normal vector angle and elevation gradient difference between adjacent supervoxel units; Performing region growing iteration on the candidate ground seed point set according to the dynamic growing rule to generate a candidate ground region set; Based on the topological connectivity of the candidate ground area set, isolated areas are eliminated by setting a maximum discontinuous span threshold to generate an optimized ground area segmentation result; The non-ground supervoxel units are reversely labeled according to the ground area segmentation result to generate ground and non-ground area segmentation results.

4. The intelligent area division method based on laser radar according to claim 1, characterized in that: The step of clustering point clouds within different distance ranges in the non-ground supervoxel unit set of the region segmentation result according to the adaptive clustering threshold to generate an obstacle cluster target set includes: Compensate and correct the point cloud density within different distance ranges based on the distribution characteristics of the lidar scan line to generate a density-corrected non-ground point cloud set; Constructing a dynamic clustering threshold function according to local point cloud density and signal-to-noise ratio parameters in the non-ground point cloud set; Performing multi-scale spatial clustering on the non-ground point cloud set by using the dynamic clustering threshold function to generate an initial obstacle cluster set; Based on the overlapping area volume ratio and density confidence parameters of the initial obstacle cluster set, low-confidence clusters are eliminated through competitive screening rules to generate an obstacle cluster target set.

5. The intelligent area division method based on laser radar according to claim 4 is characterized in that: The step of performing boundary fitting on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area includes: spatially sorting the point cloud of the obstacle cluster target set according to the topological continuity of the laser radar scan line to generate a scan line aligned obstacle point cloud sequence; According to the geometric feature distribution of the obstacle point cloud sequence, regular obstacles and irregular obstacles are fitted respectively by a multimodal boundary generation algorithm to determine an initial boundary set; Performing type annotation on the initial boundary set according to preset semantic annotation rules; Based on the spatial distribution characteristics of the annotation types, the boundary between the passable area and the obstacle area is smoothly optimized through a gridded occupancy probability model to determine the regional boundary between the passable area and the obstacle area.

6. The intelligent area division method based on laser radar according to claim 1, characterized in that: The method further comprises: generating a dynamic resource allocation weight matrix according to the supervoxel unit distribution density in the region segmentation result; Perform multi-channel iterative constraints on the expansion rate of the ground growth seed points through the dynamic resource allocation weight matrix to generate regional computing resource thresholds; The clustering threads of the non-ground supervoxel unit set are parallelized and scheduled based on the sub-region computing resource threshold, and a point cloud processing queue of the clustering threads is determined.

7. The laser radar-based intelligent area division method according to claim 6, characterized in that: The method further comprises: generating a multimodal feature fusion weight matrix based on the sub-regional computing resource threshold and the spatiotemporal synchronization data collected by the multi-source sensors; Performing spatiotemporal alignment optimization on the lidar point cloud data and other modal perception data through the multimodal feature fusion weight matrix to generate a cross-sensor calibration point cloud set; Based on the topological consistency of the cross-sensor calibration point cloud set, a global area occupancy map is constructed through a multi-resolution rasterization probability model; The global area occupancy map is incrementally updated through a dynamic Bayesian network to generate an optimized global traversable area map.

8. An intelligent area division device based on laser radar, characterized in that: The intelligent area division device based on laser radar includes: An acquisition module is used to obtain a target supervoxel unit set obtained by performing multi-scale supervoxel division on the lidar point cloud data; An expansion module, configured to generate a ground and non-ground area segmentation result by performing dynamic growth rule expansion on the ground growth seed points of the target supervoxel unit set; a clustering module, configured to cluster point clouds within different distance ranges in a non-ground supervoxel unit set of the region segmentation result according to an adaptive clustering threshold, to generate an obstacle clustering target set; The fitting module is used to perform boundary fitting on the obstacle target according to the obstacle cluster target set to determine the regional boundary between the passable area and the obstacle area.

9. An electronic device, characterized in that: Comprising a memory and a processor, wherein: The processor is configured to execute a computer program stored in the memory; When the processor executes the computer program, it implements the steps in the laser radar-based intelligent area division method described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the laser radar-based intelligent area division method described in any one of claims 1 to 7 are implemented.

Images