Networked vehicle laser radar point cloud registration acceleration method
By partitioning and density screening of the point cloud of connected vehicles and integrating virtual geometric features, the problems of calculation overhead and transmission delay of point cloud registration of connected vehicles are solved, and the accuracy of registration is improved.
Patent Information
- Application Number
- CN202411555251.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-02-11
AI Technical Summary
Point cloud registration between connected vehicles faces the problems of small overlap areas and high noise levels, which leads to increased computing overhead and reduced accuracy, and at the same time, point cloud transmission delay is serious.
By dividing point cloud areas, filtering points with higher point cloud density for registration, and fusing virtual geometric features before registration calculation to increase correspondence between point clouds.
It effectively reduces the calculation delay and transmission delay of point cloud registration, while improving the accuracy of registration.
Smart Images
Figure CN119417873A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of point cloud registration, and specifically relates to a connected vehicle laser radar point cloud registration acceleration method. Background Art
[0002] Point cloud registration facilitates accurate 3D scene reconstruction and is a fundamental element of multi-vehicle collaborative driving, helping to improve the safety of driving applications. Taking two connected vehicles as an example, the point clouds they acquire are called source point clouds and target point clouds, respectively. Through point cloud registration, we transform the target point cloud and then integrate these point clouds together. In this case, fast and accurate point cloud registration is a must for expanding the field of view of connected vehicles, which will help with subsequent tasks including collaborative perception, trajectory prediction, and path planning. Compared with high-precision maps used for on-board simultaneous localization and mapping technology, point cloud registration has a clear advantage because it avoids resource-intensive procedures such as map construction, transmission, and update. In addition, it can also show adaptability to current map-free path planning technologies in the absence of high-definition maps.
[0003] However, compared with the registration of a single vehicle, the registration of point clouds between connected vehicles faces more challenges. In LiDAR simultaneous mapping and localization technology, point clouds captured by a single vehicle at adjacent moments are usually registered. In this case, the point clouds have a large overlap area and a low noise level, making the registration process relatively simple. However, when dealing with the registration between connected cars, the conditions are the opposite, with a small overlap area and a high noise level, which poses a challenge. This increases the computational overhead and reduces the accuracy during the registration process. Although basic tools such as KPConv and MinKowski Engine have been developed to speed up the calculation of 3D points, the registration of point clouds between connected vehicles is still a computationally intensive task. In addition, the registration between connected vehicles also involves the transmission of point clouds. Considering the widely used Velodyne LiDAR HDL-64E, which can capture more than 100,000 points in a single scan, such a large amount of data will cause severe transmission delays. However, existing point cloud feature extraction methods involve extracting high-dimensional features or utilizing cross-attention mechanisms, which cannot reduce the transmission data. Summary of the invention
[0004] In view of the above technical problems, the purpose of the present invention is to provide a method for accelerating point cloud registration between connected vehicles, divide the overlapping area, further filter the points useful for registration according to the point cloud density, and fuse the virtual geometric point cloud in the registration calculation. This can effectively reduce the calculation delay and transmission delay of point cloud registration, and improve the accuracy of registration.
[0005] The technical solution of the present invention is:
[0006] A method for point cloud registration of a laser radar for a connected vehicle, comprising the following steps:
[0007] S01: The target point cloud vehicle transmits a matrix representing the point cloud distribution to the source point cloud vehicle; the source point cloud vehicle performs template matching on the matrix and its own point cloud distribution matrix to obtain an initial position.
[0008] S02: In the source point cloud vehicle, the point cloud data is divided based on the obtained initial position and point cloud density, and some features of the corresponding vehicle point cloud and the ground point cloud are retained.
[0009] S03: The source point cloud vehicle transmits the segmentation result back to the target point cloud vehicle, and the target point cloud vehicle transmits the segmented point cloud to the source point cloud vehicle for alignment calculation based on the result.
[0010] S04: Before the registration calculation, the external shape features of the vehicle are integrated into the divided point cloud using virtual geometric feature fusion to increase the correspondence for the point cloud registration.
[0011] Preferably, in step S01, the Patchworkpp algorithm is used to divide the ground point cloud into ground point cloud, non-ground low point cloud and non-ground high point cloud. The point cloud is divided into two-dimensional grids to represent the point cloud distribution. An initial positioning algorithm is designed to obtain the initial position of the vehicle using a template matching method.
[0012] Preferably, in step S02, the point cloud is further filtered according to the point cloud density of each area to obtain points that have a positive impact on the registration, while retaining the information extracted from the corresponding partial point cloud and the ground point cloud.
[0013] Preferably, in step S03, the target point cloud vehicle obtains the segmentation result of the source point cloud vehicle and transmits the required point cloud to the source point cloud vehicle.
[0014] Preferably, in step S04, before the registration calculation, the virtual geometric features of the vehicle are obtained and integrated into the registration point cloud to increase the correspondence between the registration point clouds.
[0015] Compared with the prior art, the advantages of the present invention are:
[0016] 1. In order to determine the valid points in the possible overlapping area, the present invention has developed an innovative initial positioning method based on point cloud distribution. This positioning method can determine the approximate position of the vehicle in a short time based on a small amount of information, which allows the connected vehicle to transmit less information data first.
[0017] Second, in order to reduce the transmission of non-overlapping point clouds as much as possible, the present invention can filter out points that are useless for registration and retain most of the points that have a positive effect on registration.
[0018] 3. In order to ensure the accuracy of registration, the present invention integrates virtual geometric features into these point clouds to increase the correspondence between the two point clouds, thereby improving the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0020] Figure 1 is a flow chart of the present invention;
[0021] Figure 2 is the point cloud image of this embodiment;
[0022] Figure 3 The grid divided for this embodiment;
[0023] Figure 4 The virtual geometric point cloud and the fused point cloud of this embodiment. DETAILED DESCRIPTION
[0024] The above scheme is further described below in conjunction with specific examples. It should be understood that these examples are used to illustrate the present invention and are not limited to the scope of the present invention. The implementation conditions adopted in the examples can be further adjusted according to the conditions of the specific manufacturer, and the unspecified implementation conditions are usually the conditions in conventional experiments.
[0025] In order to solve the problem in the background technology, the present invention provides a connected vehicle laser radar point cloud registration method, comprising the following steps:
[0026] S01: The target point cloud vehicle transmits a matrix representing the point cloud distribution to the source point cloud vehicle; the source point cloud vehicle performs template matching on the matrix and its own point cloud distribution matrix to obtain an initial position.
[0027] S02: In the source point cloud vehicle, the point cloud data is divided based on the obtained initial position and point cloud density, and some features of the corresponding vehicle point cloud and the ground point cloud are retained.
[0028] S03: The source point cloud vehicle transmits the segmentation result back to the target point cloud vehicle, and the target point cloud vehicle transmits the segmented point cloud to the source point cloud vehicle for alignment calculation based on the result.
[0029] S04: Before the registration calculation, the external shape features of the vehicle are integrated into the divided point cloud using virtual geometric feature fusion to increase the correspondence for the point cloud registration.
[0030] Preferably, in step S01, the Patchworkpp algorithm is used to divide the ground point cloud into ground point cloud, non-ground low point cloud and non-ground high point cloud. The point cloud is divided into two-dimensional grids to represent the point cloud distribution. An initial positioning algorithm is designed to obtain the initial position of the vehicle using a template matching method.
[0031] Preferably, in step S02, the point cloud is further filtered according to the point cloud density of each area to obtain points that have a positive impact on the registration, while retaining the information extracted from the corresponding partial point cloud and the ground point cloud.
[0032] Preferably, in step S03, the target point cloud vehicle obtains the segmentation result of the source point cloud vehicle and transmits the required point cloud to the source point cloud vehicle.
[0033] Preferably, in step S04, before the registration calculation, the virtual geometric features of the vehicle are obtained and integrated into the registration point cloud to increase the correspondence between the registration point clouds.
[0034] Specific examples Breast cancer:
[0035] The invention discloses a method for accelerating the registration of laser radar point clouds of connected vehicles. The point cloud acquired by the laser radar is divided and initially positioned according to the divided grids. Based on the approximate position, a density consistency division strategy is used to select point clouds useful for registration. The virtual geometric point cloud is integrated into the original point cloud, and then the registration is performed. Figure 1 As shown, the process of this embodiment is as follows:
[0036] (1) The ground segmentation part divides the point cloud to obtain the ground point cloud and non-ground point cloud, and divides them into grids. The target vehicle transmits the grid data to the source point cloud vehicle. In the source point cloud vehicle, the possible positions of the vehicle are generated, and template matching is performed to obtain the initial position.
[0037] (2) Based on the initial position obtained, the density consistency method is used to further filter invalid point clouds. Among them, areas with similar density are easier to find corresponding features. This method uses a specific density threshold to filter areas. In addition, the information of the corresponding vehicle point cloud and ground point cloud is also retained.
[0038] (3) The target point cloud vehicle selects the retained point cloud and transmits it to the source point cloud vehicle based on the segmentation information transmitted back by the source point cloud vehicle. Virtual geometric feature fusion is performed on the source point cloud vehicle to increase the number of correspondences, thereby improving the accuracy of point cloud registration.
[0039] In order to illustrate the framework of the present invention, this embodiment uses Figure 2The point cloud shown in the figure is divided into ground point cloud, non-ground point cloud and corresponding vehicle point cloud. The implementation steps of the connected vehicle lidar point cloud registration acceleration method are as follows:
[0040] (1) Using the PathWorkpp algorithm, such as Figure 2 The point cloud shown is divided into ground point cloud and non-ground point cloud, and each part of the point cloud is divided into grids according to the distribution, such as Figure 3 The occlusion of the vehicle will generate blank positions, and the initial position of the vehicle is obtained by searching for blank positions. In order to prevent the blank positions generated by the vehicle from being too large, which will increase the search space, a one-dimensional Gaussian distribution facing the center is deployed.
[0041] (2) Find possible corresponding points in the processed two-dimensional point cloud distribution map. The algorithm flow is as follows: Figure 4 As shown. A total of two template matchings are performed. The first time, the non-ground high point clouds of the source vehicle and the target vehicle are template matched to obtain the normalized cross-correlation value. The largest k possible corresponding positions are selected through the Top-k algorithm. The second round of template matching is performed among the k possible corresponding positions. The most likely corresponding position is selected by summing the two rounds of normalized cross-correlation values. The normalized cross-correlation value is defined as follows:
[0042]
[0043] Where M represents the matrix obtained by the point cloud distribution.
[0044] (3) Use the distribution matrix of the point cloud to design a partitioning strategy and retain point clouds in areas with similar density. s (i,j) and M s (i,j), select the corresponding grid with density ratio between 1 / and:
[0045] 1 / ≤M s (i,j) / M s (i,j)≤,(4.5)
[0046] Where represents the density threshold. This part of the point cloud is called the segmented point cloud. For different registration algorithms, the density threshold setting may be different. Observations of the vehicle that generated the target point cloud will appear in the source point cloud, and vice versa. Since the lidar cannot observe the appearance of its own vehicle, this part of the corresponding point cloud cannot establish the necessary correspondence, making it a redundant part of the registration process. However, considering the stability of the vehicle's external shape, the effective use of this point cloud will increase the number of correspondences, especially the number of correct correspondences, which will play a positive role in the accuracy of the registration. In the segmentation strategy, the corresponding point cloud is usually filtered out due to density mismatch. According to the selected possible position, the echo registration method retains this part of the corresponding point cloud by retaining a specific range of possible positions, and increases this part of the correspondence through subsequent virtual geometric feature fusion.
[0047] (4) The echo registration method needs to perform virtual geometry feature fusion, which involves introducing correspondences with the corresponding point cloud. The target vehicle transfers the virtual geometry point cloud to the source vehicle, and then extracts features as virtual geometry features. Subsequently, these features are connected with the features in the segmented point cloud. Instead of integrating the entire virtual point cloud, we trim it based on the initial position obtained to avoid adding mismatched points. The connection effectively increases the number of correspondences between the two point clouds, such as Figure 4 As shown, it places a virtual geometric feature at the center of the point cloud.
[0048] The above examples are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.
Claims
1. A connected vehicle laser radar point cloud registration acceleration method, characterized in that: The following steps are involved: S01: The target point cloud vehicle transmits a matrix representing the point cloud distribution to the source point cloud vehicle, and the source point cloud vehicle performs template matching on the matrix and its own point cloud distribution matrix to obtain the initial position; S02: In the source point cloud vehicle, the point cloud data is divided based on the obtained initial position and point cloud density, and some features of the corresponding vehicle point cloud and the ground point cloud are retained; S03: The source point cloud vehicle transmits the segmentation result back to the target point cloud vehicle, and the target point cloud vehicle transmits the segmented point cloud to the source point cloud vehicle for registration calculation according to the result; S04: Before the registration calculation, the external shape features of the vehicle are integrated into the divided point cloud using virtual geometric feature fusion to increase the correspondence for the point cloud registration.
2. The method for accelerating the registration of connected vehicle laser radar point cloud according to claim 1, characterized in that: In step S01, the Patchworkpp algorithm is used to divide the ground point cloud into ground point cloud, non-ground low point cloud and non-ground high point cloud. The above point cloud is divided into two-dimensional grids to represent the point cloud distribution, an initial positioning algorithm is designed, and the initial position of the vehicle is obtained by using the template matching method.
3. The method for accelerating the registration of connected vehicle laser radar point cloud according to claim 1, characterized in that: In step S02, the point cloud is further filtered according to the point cloud density of each area to obtain points that have a positive impact on the registration, while retaining the information extracted from the corresponding partial point cloud and the ground point cloud.
4. The method for accelerating the registration of connected vehicle laser radar point cloud according to claim 3, characterized in that: In step S03, the target point cloud vehicle obtains the segmentation result of the source point cloud vehicle and transmits the required point cloud to the source point cloud vehicle.
5. The method for accelerating the registration of connected vehicle laser radar point cloud according to claim 1, characterized in that: In the step S04, before the registration calculation, the virtual geometric features of the vehicle are obtained and integrated into the registration point cloud to increase the correspondence between the registration point clouds.