A method and system for foreground point recovery based on dynamic voxel segmentation and connectivity determination
By using dynamic voxel segmentation and connectivity discrimination, the problem of redundant ground points in the environmental perception of autonomous driving is solved, and the foreground structure is effectively preserved and noise is suppressed, thereby improving detection accuracy and efficiency.
Patent Information
- Application Number
- CN202611064497.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing ground segmentation algorithms struggle to effectively distinguish and preserve foreground structures while suppressing redundant ground points in autonomous driving environmental perception, leading to increased difficulty and decreased accuracy in subsequent detection tasks, especially in complex urban scenarios.
A method based on dynamic voxel partitioning and connectivity discrimination is adopted. Voxels are generated by adaptive partitioning in the orientation, radial and vertical dimensions, and non-ground points are recovered based on the geometric connectivity between adjacent voxels. Ground points are then identified using preset conditions.
This effectively reduced the difficulty of subsequent detection tasks, minimized interference with subsequent detection tasks, ensured the integrity and purity of key foreground points, and improved detection accuracy.
Smart Images

Figure CN123118408A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving environmental perception technology, specifically to a method and system for foreground point recovery based on dynamic voxel segmentation and connectivity discrimination. Background Technology
[0002] In the fields of autonomous driving and mobile robotics, 3D object detection based on LiDAR is a core component of environmental perception, and ground segmentation, as a crucial step in point cloud preprocessing, directly impacts the accuracy and efficiency of subsequent detection tasks. Ground segmentation methods for LiDAR point clouds have undergone years of development and can be broadly categorized into two types: traditional geometry-driven methods and deep learning methods. Traditional geometry-driven methods were the mainstream in early research, their core idea being to construct mathematical models based on the geometric characteristics of point clouds to distinguish between ground and non-ground points. Deep learning methods have become a research hotspot in recent years, automatically learning feature representations of point clouds by constructing neural network models.
[0003] Ideal ground segmentation effectively removes background points while preserving complete foreground object points, thereby optimizing the ratio of foreground to background point clouds and providing cleaner input for the detector. However, existing ground segmentation algorithms still face significant challenges in practical applications, especially when dealing with complex urban scenes. Balancing "oversegmentation" and "undersegmentation" has become a bottleneck restricting performance improvement.
[0004] In traditional geometry-driven methods, plane fitting-based approaches iteratively estimate ground plane parameters using algorithms such as RANSAC, but they are prone to undersegmentation when dealing with complex non-planar terrain. Mesh-based methods project point clouds onto a two-dimensional mesh and determine ground attributes by comparing the height of the lowest point within the mesh with a threshold, improving computational efficiency but being sensitive to mesh resolution. Patchwork++, as a representative improved algorithm, introduces adaptive local thresholding and robust ground fitting strategies, demonstrating superior performance to traditional methods on benchmark datasets such as KITTI, but it still relies on geometric assumptions and has limitations when handling steep slopes or uneven road surfaces.
[0005] In deep learning methods, early research primarily converted point clouds into bird's-eye views or distance images, using 2D convolutional neural networks for segmentation, but this suffered from information projection loss. Subsequent research shifted to point-voxel fusion, which can capture richer local and global features, but its model complexity is high, and its real-time performance is difficult to meet the needs of some application scenarios.
[0006] Compared to traditional methods, deep learning methods typically come with higher computational and time costs. Furthermore, given the lack of ground truth annotations for most 3D detection datasets, a traditional, non-learning algorithm is often used for ground point segmentation. However, traditional methods frequently suffer from "over-filtering," leading to the loss of foreground points and potentially affecting object detection accuracy. To address this issue, existing research tends to retain excessive details to meet the needs of subsequent object detection algorithms, but this often introduces a large number of noisy points. This strategy overlooks a crucial problem: excessive, unstructured details can also introduce unnecessary interference and computational burden to subsequent object detection algorithms.
[0007] Therefore, there is an urgent need for a foreground point recovery method and system based on dynamic voxel segmentation and connectivity discrimination, which can effectively distinguish and preserve the foreground structure while suppressing the recovery of redundant ground points, thereby reducing the difficulty of subsequent detection tasks and minimizing interference with subsequent detection tasks while ensuring the integrity of key foreground points. Summary of the Invention
[0008] One of the objectives of this invention is to provide a foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination, which can effectively distinguish and preserve foreground structures while suppressing the recovery of redundant ground points, thereby reducing the difficulty of subsequent detection tasks and minimizing interference with subsequent detection tasks while ensuring the integrity of key foreground points.
[0009] The basic solution provided by this invention is a foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination, which includes the following: S1. Model the non-uniform point cloud structure and generate voxels by adaptive partitioning of the orientation, radial and vertical dimensions; S2. Based on the geometric connectivity between adjacent voxels, ground points in voxels that meet the corresponding preset conditions are restored to non-ground points; where the preset conditions are the relationship between the radial distance of the ground points and the maximum radial distance of the non-ground points in the voxels.
[0010] Furthermore, S1 includes: S101. Convert the point cloud data from the Cartesian coordinate system to the cylindrical coordinate system for representation, and obtain the cylindrical coordinates of the point cloud data; S102. Based on the cylindrical coordinate system, the voxel space is divided into regions along the three dimensions of orientation, radial direction and vertical direction.
[0011] Furthermore, based on azimuth dimensions, the region is divided: in the bird's-eye view, the point cloud is projected onto the horizontal plane and divided equally along the azimuth direction. Each sector is a separate sector, the size of which is determined by the hyperparameter angular resolution. The chosen strategy for directional area division is as follows:
[0012] in, Point The directional area division number, This indicates the angular resolution of the azimuth region division.
[0013] Furthermore, based on the radial dimension, the regions are divided: the LiDAR installed on the autonomous vehicle projects the point cloud onto the horizontal plane from a bird's-eye view, and divides it into N independent regions along the radial direction. The radial region division strategy is as follows:
[0014] in, Point Radial region division numbering, Indicates the angular resolution of the radial region division. Indicates the horizontal height of the lidar. This indicates the radial distance that the lidar filters.
[0015] Furthermore, based on the vertical dimension, the region is divided: according to the vertical height of the point cloud, it is divided along the vertical direction into... Each region is independent, and the vertical region division strategy is as follows:
[0016] in, Point Vertical region division numbering, Indicates the angular resolution of the vertical region division. This represents the vertical angle reference parameter of the lidar.
[0017] Furthermore, S2 includes: S201. Determine the state of each voxel based on the type of point cloud inside the voxel; S202. Based on the state of the voxels, determine whether the ground points in voxels containing only ground points and voxels containing both ground points and non-ground points meet the corresponding preset conditions. If they do, restore them to non-ground points.
[0018] Furthermore, the states include: empty voxels, containing only non-ground points, containing only ground points, and containing both ground and non-ground points; the voxels corresponding to each state are: empty voxels, pure non-ground point voxels, pure ground point voxels, and mixed voxels; S202 includes: If the voxel is an empty voxel, no processing is performed; If the voxel is a pure non-ground point voxel, then record the maximum and minimum values of its radial distance; If a voxel is a pure ground point voxel, its attributes are determined by checking its adjacent preceding voxels: if the adjacent voxel is empty or is also a pure ground point voxel, the points in the current voxel are identified as ground points and are not restored; otherwise, the points in the voxel will be judged according to the first preset condition, and the points that meet the first preset condition will be restored as non-ground points. After restoration, the maximum and minimum radial distance values will be updated once. If the voxel is a mixed voxel, first calculate the extreme values of the radial distance of the non-ground points within the voxel, then determine whether the ground points within the voxel meet the second preset condition, restore the ground points that meet the second preset condition to non-ground points, and leave the remaining points unchanged. Based on the determined point cloud, recalculate and update the maximum and minimum values of the radial distance of the non-ground points within the voxel.
[0019] Furthermore, the first preset condition is:
[0020] in, Point radial distance, The azimuth, radial, and perpendicular numbers are respectively The maximum radial distance of non-ground points in a voxel. Voxel representation Minimum radial distance between points on the ground in Central Africa It is a hyperparameter, representing the allowable error.
[0021] Furthermore, the second preset condition is: .
[0022] The second objective of this invention is to provide a foreground point recovery system based on dynamic voxel partitioning and connectivity discrimination, which can effectively distinguish and retain foreground structures while suppressing the recovery of redundant ground points, thereby reducing the difficulty of subsequent detection tasks and minimizing interference with subsequent detection tasks while ensuring the integrity of key foreground points.
[0023] This invention provides a second basic solution: a foreground point recovery system based on dynamic voxel partitioning and connectivity discrimination, comprising: a dynamic voxel partitioning structure module and a connectivity discrimination and recovery module; The dynamic voxel partitioning structure submodule is used to model non-uniform point cloud structures and generate voxels through adaptive partitioning of orientation, radial and vertical dimensions. The connectivity discrimination and restoration module is used to restore ground points in voxels that meet corresponding preset conditions to non-ground points based on the geometric connectivity between adjacent voxels; wherein the preset conditions are the relationship between the radial distance of the ground point and the maximum radial distance of the non-ground points in the voxel.
[0024] Beneficial effects: This solution models the non-uniform point cloud structure by adaptively dividing the azimuth, radial and vertical dimensions. This not only effectively reduces the huge computational overhead of processing the original LiDAR point cloud data in the form of disordered Cartesian coordinates, but also adapts well to the non-uniform distribution of point clouds, solving the obvious adaptability defects of the traditional fixed-size voxel division strategy in the scenario of LiDAR usually installed on the top of the vehicle to perceive the surrounding environment. Based on the geometric connectivity between adjacent voxels, ground points within a voxel that meet corresponding preset conditions are restored as non-ground points. These preset conditions relate the radial distance between the ground point and the maximum radial distance between non-ground points within the voxel. This effectively distinguishes and preserves the foreground structure while suppressing the restoration of redundant ground points, thus reducing the difficulty of subsequent detection tasks. It minimizes interference with subsequent detection tasks while ensuring the integrity of key foreground points. This effectively solves the problems of traditional methods losing foreground details due to "over-segmentation" and strategies such as Attached Points Restoring (APR) introducing noise due to "over-restoration."
[0025] In summary, this scheme can effectively distinguish and preserve the foreground structure while suppressing the recovery of redundant ground points, thereby reducing the difficulty of subsequent detection tasks and minimizing interference with subsequent detection tasks while ensuring the integrity of key foreground points. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 2 This is a schematic diagram of the orientation region division in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention; Figure 3 This is a schematic diagram of the installation of LiDAR in an autonomous vehicle in an embodiment of the foreground point recovery method based on dynamic voxel segmentation and connectivity discrimination of the present invention. Figure 4 This is a schematic diagram of radial region division in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention; Figure 5 This is a schematic diagram of vertical region division in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention; Figure 6This is a schematic diagram of the voxel state in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 1 ; Figure 7 This is a schematic diagram of the voxel state in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 2 ; Figure 8 This is a schematic diagram of the voxel state in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 3 ; Figure 9 This is a schematic diagram of the voxel state in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 4 ; Figure 10 This is an overall image of the point cloud processing result of Patchwork++ combined with APR in an embodiment of the foreground point restoration method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 11 This is a detailed illustration of the point cloud processing results using Patchwork++ combined with APR in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 12 This is an overall image of the point cloud processing result of Patchwork++ combined with PRM in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 13 This is a detailed illustration of the point cloud processing results of Patchwork++ combined with PRM in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Figure 14 This is a flowchart illustrating the process of using Patchwork++ combined with PRM in an embodiment of the foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination of the present invention. Detailed Implementation
[0027] The following detailed description illustrates the specific implementation method: Example 1 This embodiment provides a foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination, as shown in the attached figure. Figure 1 As shown, it includes the following: S1. The non-uniform point cloud structure is modeled and voxels are generated through adaptive partitioning of azimuth, radial and vertical dimensions; the modeling generates region partitioning results of different dimensions, i.e., voxels (dynamic voxels). S101. Convert the point cloud data from the Cartesian coordinate system to the cylindrical coordinate system for representation, and obtain the cylindrical coordinates of the point cloud data; Specifically, the conversion formula is:
[0028] in, Indicates radial distance. Indicates azimuth. Let be the coordinates of the point in the Cartesian coordinate system.
[0029] S102. Based on the three dimensions of cylindrical coordinates—azimuth, radial distance, and vertical height—the voxel space is divided into regions, so that each voxel contains the corresponding point cloud data.
[0030] S102. Based on the cylindrical coordinate system, the voxel space is divided into regions along the three dimensions of orientation, radial direction and vertical direction. Each voxel is a three-dimensional (orientation, radial, vertical) space (i.e., voxel space), shaped like an arc. Each voxel (arc) contains points. Based on the state of the points within its interior, voxels (arcs) are classified into four categories: an empty voxel if it contains no points; a pure non-ground voxel if all points are non-ground points; a pure ground voxel if all points are ground points; and a mixed voxel if it contains both ground and non-ground points. Specifically, the region is divided according to azimuth: in a bird's-eye view, the point cloud is projected onto a horizontal plane and divided equally along the azimuth direction. Each sector is a separate sector, the size of which is determined by the hyperparameter angular resolution. Decision, such as Figure 2 As shown; The strategy for dividing the directional area is as follows:
[0031] in, Point The directional area division number, This indicates the angular resolution of the azimuth region division.
[0032] Based on the radial dimension, the region is divided: LiDAR installed on autonomous vehicles, such as... Figure 3 As shown, also based on a bird's-eye view, the point cloud is projected onto a horizontal plane and divided into N independent regions along the radial direction, as follows: Figure 4 As shown; The radial region partitioning strategy is as follows:
[0033] in, Point Radial region division numbering, Indicates the angular resolution of the radial region division. Indicates the horizontal height of the lidar. Indicates the radial distance of the lidar filter. This indicates rounding down. This invention uses this formula to obtain the region number, which is an integer. This invention divides the interval [0,1) into region number 0, so 0.1 and 0.9 should both be in region 0, hence the use of rounding down.
[0034] Based on the vertical dimension, the region is divided: according to the vertical height of the point cloud, it is divided along the vertical direction into... An independent region, such as Figure 5 As shown; The vertical region division strategy is as follows:
[0035] in, Point Vertical region division numbering, Indicates the angular resolution of the vertical region division. This represents the vertical angle reference parameter of the lidar.
[0036] On the one hand, since LiDAR point clouds exist in disordered Cartesian coordinates, directly processing such raw data incurs enormous computational overhead. On the other hand, the point cloud distribution exhibits significant non-uniformity, resulting in high point density near the sensor and sparser distribution further away. Given that LiDAR is typically mounted on the roof of a vehicle to perceive its surroundings, traditional fixed-size voxel partitioning strategies have significant adaptability limitations in this scenario, leading to an overabundance of samples in nearby voxels and a scarcity of samples in distant voxels. Therefore, this solution proposes the aforementioned dynamic voxel partitioning method based on point cloud distribution. It primarily models the non-uniform point cloud structure through adaptive partitioning in azimuth, radial, and vertical dimensions. This not only effectively reduces the enormous computational overhead of processing raw LiDAR point cloud data in disordered Cartesian coordinates but also adapts well to the non-uniform distribution of point clouds, solving the significant adaptability limitations of traditional fixed-size voxel partitioning strategies in scenarios where LiDAR is typically mounted on the roof of a vehicle to perceive its surroundings.
[0037] S2. Based on the geometric connectivity between adjacent voxels, ground points in voxels that meet the corresponding preset conditions are restored to non-ground points. That is, the PRM algorithm is used to process them sequentially according to the voxel numbering order to more accurately restore foreground points (non-ground points), intelligently restore key foreground points that have been missegmented, and effectively suppress noise interference. S201. Determine the state of each voxel based on the type of point cloud within the voxel; the states include: empty voxel, containing only non-ground points, containing only ground points, and containing both ground and non-ground points; the voxels corresponding to each state are: empty voxel, pure non-ground point voxel, pure ground point voxel, and mixed voxel; for example... Figure 6-9 As shown, red dots represent non-ground points retained after coarse segmentation, while green dots represent points initially identified as ground. S202. Based on the state of the voxels, determine whether the ground points in voxels containing only ground points and voxels containing both ground points and non-ground points meet the corresponding preset conditions. If they do, restore them to non-ground points.
[0038] Specifically, in this embodiment, the data is processed sequentially according to the voxel numbering order: If the voxel is an empty voxel, that is, the state is an empty voxel, then no processing is performed; If the voxel is a pure non-ground point voxel, that is, the state contains only non-ground points, then record the maximum and minimum values of its radial distance; If the voxel is a pure ground point voxel, i.e., the state contains only ground points, then by checking its adjacent preceding voxels (i.e., the azimuth, radial, and vertical numbers are respectively...) The voxels are used to determine their attributes: if the adjacent voxel is empty or is also a pure ground point voxel, the point in the current voxel is identified as a ground point and is not restored; otherwise, the point in the voxel will be judged according to the first preset condition, and the point that meets the first preset condition will be restored as a non-ground point. After the restoration is completed, the maximum and minimum radial distance values are updated once. The first presupposition condition is:
[0039] in, Point radial distance, The azimuth, radial, and perpendicular numbers are respectively The maximum radial distance of non-ground points in a voxel. Voxel representation Minimum radial distance between points on the ground in Central Africa It is a hyperparameter, representing the allowable error.
[0040] If the voxel is a mixed voxel, i.e., the state includes ground points and non-ground points, then the extreme values of the radial distance of the non-ground points within the voxel are first calculated. The ground points within the voxel are then judged to see if they meet the second preset condition. Ground points that meet the second preset condition are restored to non-ground points, while the remaining points remain unchanged. The maximum and minimum values of the radial distance of the non-ground points within the voxel are recalculated and updated based on the judged point cloud. In this embodiment, the background point specifically refers to the ground point, and the foreground point specifically refers to the set of all non-ground point clouds other than the background point (ground point) mentioned above, which specifically includes, but is not limited to, dynamic targets and static obstacles.
[0041] The second presupposition condition is: .
[0042] The foreground point restoration based on dynamic voxel segmentation and connectivity discrimination in this scheme can effectively distinguish and preserve the foreground structure, while suppressing the restoration of redundant ground points, thereby reducing the difficulty of subsequent detection tasks. While ensuring the integrity of key foreground points, it minimizes interference with subsequent detection tasks. This effectively solves the problems of traditional methods losing foreground details due to "over-segmentation" and strategies such as Attached Points Restoring (APR) introducing noise due to "over-restoration".
[0043] To analyze the effectiveness of this proposed solution, specifically, to verify the effectiveness of the proposed PRM method, the KITTI dataset was selected as the experimental data. Patchwork++ was used as the ground coarse segmentation algorithm, and the IA-SSD detection model was specifically retrained based on the point cloud distribution after ground removal to adapt to the characteristics of the processed data. The new method was then compared with the existing state-of-the-art method APR. The specific point cloud processing results are as follows: Figure 10-13 As shown.
[0044] in, Figure 10 and Figure 11 This is a diagram showing the overall structure and details of Patchwork++ combined with the APR method. Figure 12 and Figure 13 The diagram shows the corresponding results of Patchwork++ combined with the PRM method of this scheme; the red point cloud in the diagram represents the non-ground points obtained by Patchwork++ segmentation, the green point cloud represents the ground points, and the blue point cloud represents the non-ground points recovered from the ground points by the PRM method and the APR method.
[0045] The comparison results show that Patchwork++ exhibits significant oversegmentation, leading to the loss of object details. While Patchwork++ combined with the APR method recovers some point cloud data, it introduces a large number of redundant ground points. In contrast, the proposed method effectively removes ground points while preserving the detailed structure of foreground objects to the greatest extent, achieving superior segmentation and recovery results.
[0046] In addition, to provide detailed detection results for different categories in the observation dataset, the average precision (API) was used. ) and its mean ( ) is used as the main performance evaluation indicator; among which, The calculation is shown in the formula:
[0047] in, Indicates the number of target categories. Indicates the first Average precision of the class.
[0048] Tables 1-3 show the detailed detection results for the three key categories, Car, Pedestrian, and Cyclist, on the KITTI dataset; among them, According to official evaluation standards, targets are categorized into three levels—Easy, Moderate, and Hard—based on their degree of occlusion / truncation. Three levels under this category The average value.
[0049] Table 1: Experimental Results of KITTI Dataset Car (IoU=0.7)
[0050] Table 2: Experimental Results of KITTI Dataset Ped (IoU=0.5)
[0051] Table 3: Experimental Results of the KITTI Dataset Cyclist (IoU=0.5)
[0052] As shown in Tables 1-3, the IA-SSD+PRM model outperforms the original IA-SSD model across all three categories and all difficulty levels, with mAP improvements of 0.51% (Car), 0.39% (Pedestrian), and 3.03% (Cyclist), respectively. This indicates that the proposed foreground point recovery method can effectively enhance the perception capabilities of downstream detectors by providing a more complete and cleaner foreground point cloud, with particularly significant performance improvements in the Cyclist category, which has high requirements for point cloud detail.
[0053] While the improvement achieved by this scheme compared to the IA-SSD+APR model in the Car and Pedestrian categories is small, it is stable and outperforms in the Moderate and Hard categories. This indicates that, compared to the potential "over-restoration" tendency of APR, the voxel connectivity-based discrimination mechanism adopted by PRM achieves a better balance between restoring key foreground points and suppressing distant sparse noise points, thus providing the detector with higher-quality input point clouds, especially in complex scenes where the target is partially occluded (Moderate) or severely occluded (Hard).
[0054] This scheme demonstrates stable performance improvements in the Car category (IoU=0.7), which requires extremely high detection accuracy, the Pedestrian category with sparse point clouds, and the Cyclist category, which is sensitive to shape and has complex point cloud structures. This verifies that the proposed dynamic voxel partitioning strategy can adapt well to the non-uniform distribution of point clouds, and the connectivity-based recovery mechanism is robust to foreground objects of different scales and shapes, without relying on specific prior knowledge of objects.
[0055] In this scheme, the dynamic voxel segmentation strategy can adapt well to the non-uniform distribution of point clouds, and the connectivity-based recovery mechanism is robust to foreground objects of different scales and shapes, without relying on specific prior knowledge of objects. While effectively removing ground points, it preserves the detailed structure of foreground objects to the greatest extent, achieving better segmentation and recovery results. It can effectively distinguish and preserve foreground structures while suppressing the recovery of redundant ground points, thereby reducing the difficulty of subsequent detection tasks. While ensuring the integrity of key foreground points, it minimizes interference with subsequent detection tasks. By providing clean and complete point cloud input, it effectively enhances the perception capabilities of downstream detectors.
[0056] This embodiment also provides a foreground point recovery system based on dynamic voxel partitioning and connectivity discrimination, including: a dynamic voxel partitioning structure module and a connectivity discrimination and recovery module; The Dynamic Voxel Partitioning Sub-module (DVPS) is used to model non-uniform point cloud structures and generate voxels through adaptive partitioning in the orientation, radial, and vertical dimensions. The Connectivity Discrimination & Recovery Sub-module (CDRS) is used to restore ground points in voxels that meet corresponding preset conditions to non-ground points based on the geometric connectivity between adjacent voxels; the preset conditions are the relationship between the radial distance of the ground point and the maximum radial distance of the non-ground points in the voxel.
[0057] The specific implementation process is as follows: Figure 14 As shown, a foreground point recovery system based on dynamic voxel segmentation and connectivity discrimination is used to execute the foreground point recovery method based on dynamic voxel segmentation and connectivity discrimination, referred to as the foreground recovery method (PRM). Considering the strict real-time requirements of ground segmentation as a preprocessing step, and based on research, the Patchwork++ algorithm exhibits superior segmentation accuracy and running efficiency compared to traditional methods. Therefore, the Patchwork++ algorithm is used for initial ground removal. The result after Patchwork++ processing is ground points and non-ground points, which is... Figure 14 The ground is coarsely segmented. Based on this, to compensate for information loss caused by over-segmentation, a foreground point recovery module (PRM) is further introduced to refine the point cloud.
[0058] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A foreground point recovery method based on dynamic voxel partitioning and connectivity determination, characterized in that, Includes the following: S1. By adaptively dividing the orientation, radial and vertical dimensions, the non-uniform point cloud structure is modeled and voxels are generated. S2. Based on the geometric connectivity between adjacent voxels, restore the ground points in the voxels that meet the corresponding preset conditions to non-ground points. The preset condition is the relationship between the radial distance of the ground point and the maximum radial distance of the non-ground point in the voxel.
2. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 1, characterized in that, S1 includes: S101. Convert the point cloud data from the Cartesian coordinate system to the cylindrical coordinate system for representation, and obtain the cylindrical coordinates of the point cloud data; S102. Based on the cylindrical coordinate system, the voxel space is divided into regions along the three dimensions of orientation, radial direction and vertical direction.
3. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 2, characterized in that, Based on azimuth, the region is divided: In a bird's-eye view, the point cloud is projected onto a horizontal plane and divided equally along the azimuth direction. Each sector is a separate sector, the size of which is determined by the hyperparameter angular resolution. The chosen strategy for dividing the geographical area is as follows: in, Point The directional area division number, Indicates the angular resolution of the azimuth region division. This indicates rounding down to the nearest integer.
4. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 2, characterized in that, Based on the radial dimension, the region is divided: The LiDAR installed on the autonomous vehicle projects the point cloud onto the horizontal plane from a bird's-eye view and divides it into N independent regions along the radial direction. The radial region division strategy is as follows: in, Point Radial region division numbering, Indicates the angular resolution of the radial region division. Indicates the horizontal height of the lidar. Indicates the radial distance of the lidar filter. This indicates rounding down to the nearest integer.
5. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 2, characterized in that, Based on the vertical dimension, the region is divided: according to the vertical height of the point cloud, it is divided along the vertical direction into... Each independent region is divided using the following vertical region division strategy: in, Point Vertical region division numbering, Indicates the angular resolution of the vertical region division. This represents the vertical angle reference parameter of the lidar. This indicates rounding down to the nearest integer.
6. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 1, characterized in that, S2 includes: S201. Determine the state of each voxel based on the type of point cloud inside the voxel; S202. Based on the state of the voxels, determine whether the ground points in voxels containing only ground points and voxels containing both ground points and non-ground points meet the corresponding preset conditions. If they do, restore them to non-ground points.
7. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 6, characterized in that, The states include: empty voxels, containing only non-ground points, containing only ground points, and containing both ground and non-ground points; the voxels corresponding to each state are: empty voxels, pure non-ground point voxels, pure ground point voxels, and mixed voxels. S202 includes: If the voxel is an empty voxel, no processing is performed; If the voxel is a pure non-ground point voxel, then record the maximum and minimum values of its radial distance; If a voxel is a pure ground point voxel, its attributes are determined by checking its adjacent preceding voxels: if the adjacent voxel is empty or is also a pure ground point voxel, the points in the current voxel are identified as ground points and are not restored; otherwise, the points in the voxel will be judged according to the first preset condition, and the points that meet the first preset condition will be restored as non-ground points. After restoration, the maximum and minimum radial distance values will be updated once. If the voxel is a mixed voxel, first calculate the extreme values of the radial distance of the non-ground points within the voxel, then determine whether the ground points within the voxel meet the second preset condition, restore the ground points that meet the second preset condition to non-ground points, and leave the remaining points unchanged. Based on the determined point cloud, recalculate and update the maximum and minimum values of the radial distance of the non-ground points within the voxel.
8. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 7, characterized in that, The first preset condition is: in, Point radial distance, The azimuth, radial, and perpendicular numbers are respectively The maximum radial distance of non-ground points in a voxel. Voxel representation Minimum radial distance between points on the ground in Central Africa It is a hyperparameter, representing the allowable error.
9. The foreground point recovery method based on dynamic voxel partitioning and connectivity discrimination according to claim 8, characterized in that, The second preset condition is: 。 10. A foreground point recovery system based on dynamic voxel partitioning and connectivity discrimination, characterized in that, include: Dynamic voxel partitioning structure module and connectivity discrimination and recovery module; The dynamic voxel partitioning structure module is used to model non-uniform point cloud structures and generate voxels through adaptive partitioning in the orientation, radial and vertical dimensions. The connectivity discrimination and restoration module is used to restore ground points in voxels that meet the corresponding preset conditions to non-ground points based on the geometric connectivity between adjacent voxels. The preset condition is the relationship between the radial distance of the ground point and the maximum radial distance of the non-ground point in the voxel.
