A roadside multi-lidar registration method based on active features and mobile relay

Abstract

The application discloses a roadside multi-laser radar registration method based on active features and mobile relays, first, all laser radar sites and active geometric feature shape markers of a public area are arranged, second, multiple laser radar point clouds and GPS timestamp data are taken as input, and after preprocessing, a to-be-registered area containing active features is demarcated; coarse registration parameters between source point clouds and target point clouds are obtained through regional normal distribution transformation, active geometric feature space representation equations between the two to-be-registered point clouds are fitted, and strong correlation features between the two laser radar coordinates, namely public key points, are solved; with the aid of key feature point pairs, accurate pose parameters between the two laser radar points are solved based on a nearest point matching algorithm and singular value decomposition, so that the same registration operation is repeated for each adjacent laser radar point cloud, pose transformation parameters of adjacent roadside points are obtained, and finally, the whole tunnel scene is integrated in a high-precision global coordinate, and the precision is higher.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation engineering technology, and mainly relates to a roadside multi-liquid radar registration method based on active features and mobile relay. Background Technology

[0002] The inability to obtain accurate global positioning and traffic scene perception information has led to serious threats to the lives of drivers and passengers due to road traffic safety management issues. To overcome this problem, most solutions focus on improving the vehicle's own perception capabilities, assisting drivers in perceiving the surrounding environment by installing onboard cameras or LiDAR sensors, which has reduced traffic accidents to some extent. However, onboard computing devices and sensors consume a large amount of vehicle resources, and due to the limitations of sensor installation height, the field of view perceived by the vehicle itself is very limited. Frequent mutual obstruction between vehicles leads to high-frequency blind spots, making it difficult to effectively ensure road safety and respond to emergencies. The inaccuracy and lag in positioning in obstructed scenarios such as tunnels further delay emergency rescue and handling.

[0003] Against this backdrop, vehicle-road cooperative intelligent sensing technology, which integrates sensing sensors deployed on roadside units, offers a new solution for improving traffic safety. Roadside sensors, with their relatively high installation position, gain a wider field of view and sensing range, thus significantly reducing blind spots caused by vehicle obstruction. By leveraging vehicle-to-infrastructure communication, the acquired real-time vehicle location information and regional road traffic conditions are broadcast to surrounding vehicles, promptly reminding drivers to drive cautiously and effectively reducing the risk of traffic accidents.

[0004] LiDAR, as the primary sensor for roadside perception and localization, can provide accurate 3D information. However, considering that the detection range of a single roadside LiDAR can only cover a very limited area, to achieve effective vehicle perception and assisted localization in tunnels, multiple LiDARs at different locations and distances along both sides of the road are needed to cover the entire tunnel traffic scenario. However, the point cloud data generated by each LiDAR is represented independently in a local coordinate system with itself as the origin. Therefore, the primary condition for achieving global perception is multi-LiDAR extrinsic parameter calibration, that is, fusing point cloud data from multiple LiDARs and representing them in a unified global coordinate system, i.e., point cloud registration.

[0005] Coordinate integration of multiple laser point clouds relies on pairwise point cloud registration, specifically solving for the three attitude angle parameters R(α,β,γ) and three displacement parameters T(△x,△y,△z) of the source point cloud Q relative to the target point cloud P. Commonly used point cloud registration methods include the Normal Distributions Transform (NDT) algorithm based on statistical probability and the Iterative Closest Point (ICP) method based on global search. The NDT algorithm estimates the distribution of the source point cloud by calculating the probability density function of the rasterized point cloud and then matches it on the target point cloud accordingly. The NDT algorithm is sensitive to raster scale; a large raster size leads to decreased accuracy, while a small raster size results in excessive memory usage, making it difficult to balance computation time and accuracy. The ICP algorithm uses the least squares method as its core computational method. It directly searches for the nearest point in the target point cloud as the corresponding point pair based on the source point cloud, solves for the transformation matrix, and obtains the final result by setting the number of iterations or a loss threshold. The ICP method based on global point-pair search does not include local shape information, thus requiring high initial parameters and having relatively low computational efficiency. Furthermore, due to the homogeneity of the geometric structure in tunnel scenes, point clouds are highly homogeneous, making it difficult to apply general point cloud registration methods to the registration of roadside lidar. In addition, considering that point clouds gradually become sparser with increasing distance, directly matching two roadside lidars within the effective sensing range can cause the iterative algorithm to fail to converge, resulting in huge errors or matching failures. Summary of the Invention

[0006] This invention addresses the problems of traditional point cloud registration methods in the prior art, such as the inability to match sparse points over long distances and the inability to register homogeneous point clouds in tunnel scenes. It provides a roadside multi-LiDAR registration method based on active features and mobile relays. The method first places active geometric feature shape markers for all LiDAR locations and common areas. Then, using multiple LiDAR point clouds and GPS timestamp data as input, and after preprocessing such as filtering, denoising, and plane fitting, a registration region containing active features is defined. Next, a designed regional normal distribution transformation is used to obtain coarse registration parameters between the source and target point clouds. The active geometric feature space representation equation between the two point clouds to be registered is fitted, and the strong correlation features between the two LiDAR coordinates, i.e., common key points, are solved. Using key feature point pairs, the singular value decomposition algorithm based on nearest-neighbor matching is applied to solve for the precise pose parameters between the two LiDAR locations. Finally, the same registration operation is repeated for each adjacent LiDAR point cloud to obtain the pose transformation parameters of adjacent roadside points, thereby achieving high-precision global coordinate integration of the entire tunnel scene. The method of this invention not only effectively improves the convergence time, but also obtains high-precision point-to-point matching rotation and translation matrix parameters, effectively overcoming the problem of difficult registration of long-distance roadside lidar within the perception range, and better serving accurate perception and assisted positioning in tunnel scenarios such as satellite rejection and point cloud homogenization.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a roadside multi-liquid radar registration method based on active features and mobile relays, comprising the following steps:

[0008] S1, Deploy roadside locations: The roadside locations include lidar locations and geometric active features, wherein the lidar locations include a main lidar P, a mobile relay lidar Q, and a lidar M to be registered.

[0009] S2, Continuous sampling point set: Collect the roadside lidar point sets at the main lidar P, mobile relay lidar Q and lidar M to be registered, merge the same points respectively, and obtain the main roadside target domain point set P, the relay source domain point set Q and the roadside source domain point set M.

[0010] S3, Point Cloud Set Preprocessing: Perform preprocessing operations on the point sets P, M, Q obtained in step S2. The preprocessing includes at least removing device noise, extracting irrelevant point clouds, plane fitting, and coordinate correction.

[0011] S4, Select active feature region: Manually select a point cloud region containing actively deployed geometric feature identifiers. The selection criterion is approximately 2 meters before and after the active feature in the x direction under the lidar coordinate system, with no restrictions in the y and z directions.

[0012] S5, NDT coarse matching initial pose: The NDT algorithm is used to perform coarse registration on the region selected in step S4 to determine the key point region for precise registration;

[0013] S6, Fitting Active Keypoints: Based on the three common geometric features deployed under the two point clouds to be registered, solve the spatial representation equations of each geometric feature under the source point cloud and the target point cloud respectively. Taking a cylinder with radius r and height h as an example, its spatial representation equation is:

[0014]

[0015] Where (A, B) are the coordinates of the center of the circle on the ground (xoy plane); according to the coordinate correction result in step S3, the ground point cloud height is 0. Based on the least squares method to limit the error, the three-dimensional spatial points of the centroid and the upper and lower bottom centers of the active feature marker are calculated as follows: Similarly, find the spatial coordinates of the centroid and the center of the upper and lower bases of the other two geometric features (A,B,0) and (A,B,h).

[0016] S7: ICP fine matching based on SVD decomposition: Point cloud fine registration based on point-to-point calculation of singular value decomposition;

[0017] S8: Pose transfer and positioning broadcast: Repeat steps S4-S7, iterate through all LiDAR points, complete pose matching, and calculate the relative pose relationship of all roadside LiDAR points in sequence through the rotation and translation relationship between two adjacent LiDARs.

[0018] As an improvement of the present invention, the point cloud set preprocessing in step S3 specifically includes:

[0019] S31: Using a point set as input, Gaussian filtering is used to remove device noise;

[0020] S32: Use KD-Tree to find all point sets within a filtering radius r, calculate the standard deviation of point cloud curvature and density based on statistical filtering, and set a threshold to remove irrelevant points;

[0021] S33: Fit the ground equation using the least squares method to obtain the parameters of the ground equation z=ax+by+c, and remove the ground points from each lidar point set;

[0022] S34: Fit the height of the point cloud on the z-axis using the ground equation, and perform coordinate transformation on each point set based on the rotation vector obtained in step S33 and the translation vector obtained in this step to perform coordinate correction.

[0023] As another improvement of the present invention, in the Gaussian filtering of step S31, the point set P(x,y,z) is used as input, G is a Gaussian function, the window radius is r, and the standard deviation of the Gaussian function is σ. The filtering formula is as follows:

[0024]

[0025] Where P′ is the filtered point cloud set, w is the total weight, N is the number of points within the window radius r, and p is the coordinate of the current point. i These are the coordinates of each point within the window.

[0026] As another improvement of the present invention, step S5 NDT coarse matching initial pose specifically includes the following steps:

[0027] S51: Divide the source point cloud into three-dimensional grids and calculate the mean and covariance of each grid point;

[0028] S52: Rotate and translate the source point cloud according to the initialized pose transformation parameters;

[0029] S53: Integrate the rotated and translated points with the target point in the same coordinate system, and calculate the probability that the new point cloud obtained according to the normal distribution falls within the reference point cloud. The overlap function f between the source point cloud and the reference point cloud after the T-transform is:

[0030]

[0031] Where q i The source point cloud is μ, the centroid of the point cloud is cov, and the point cloud covariance is covariance. Then, the NDT registration score is obtained by summing the probability densities calculated from all the grids.

[0032] S54: The objective function is optimized using the Newton-Raphson optimization algorithm to obtain the transformation parameter τ with the highest registration score.

[0033] As another improvement of the present invention, step S7, which involves fine registration of the point cloud based on point-to-point calculation using singular value decomposition, specifically includes:

[0034] S71: Based on the three fitted active geometric feature points obtained in step S6, the target point cloud P can be obtained. * {p1,p2,...,p n}, Q * {q1,q2,...,q n Where n = 9, the centroid and covariance are calculated based on the point set;

[0035] S72: Perform SVD singular value decomposition on the covariance to obtain the rotation matrix R*, thereby deriving the translation vector t* and obtaining the transformation matrix T. * .

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) This invention proposes a registration method for multiple roadside lidar point clouds in tunnel scenarios. By using relay lidar to reduce the distribution difference of point clouds of different densities on the same object, the target point cloud and the source point cloud can obtain richer point cloud coverage density.

[0038] (2) In view of the phenomenon that it is difficult for point cloud spatial points to completely overlap, the present invention uses regular active features as the common feature region of the source point cloud and the target point cloud. By fitting the active geometric feature spatial expression, the centroid and the upper and lower bottom centers are obtained as strong correlation features for point cloud registration.

[0039] (3) This invention first uses normal distribution transformation to perform coarse matching of point clouds, then fits the spatial equation of active geometric features to obtain matching key points, and uses the least squares method to implement point-to-point precise matching. This method can not only achieve the purpose of rapid convergence, but also help improve the registration accuracy. Through the transfer of the transformation matrix between relay lidars, the long-distance point cloud matching work within the effective sensing range is realized. Attached Figure Description

[0040] Figure 1 This is a flowchart of the steps of the roadside multi-liquid radar registration method based on active features and mobile relay of the present invention;

[0041] Figure 2 This is a top-view schematic diagram of the selection of the target point cloud and source point cloud regions in step S4 of the method of the present invention;

[0042] Figure 3 This is a flowchart of step S5 of the method of the present invention, which involves normal distribution transformation to register the point cloud blocks in the region. Detailed Implementation

[0043] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0044] Example 1

[0045] A roadside multi-LiDAR registration method based on active features and mobile relay addresses the technical challenge of designing a point cloud registration method for long distances and addressing severe homogenization issues. This method improves the accuracy of long-distance point cloud registration and desensitizes homogenization issues such as tunnels, avoiding distribution differences of point clouds with different densities on the same object, and achieving accurate positioning of roadside equipment with full coverage throughout the entire tunnel. Figure 1 As shown, the method of the present invention specifically includes the following steps:

[0046] Step S1: Deploy roadside sites, which include lidar sites and geometric active features.

[0047] Deploy the main lidar P, the mobile relay lidar Q, and the lidar M to be registered; deploy three cylinders with radii r and heights h, h, and 2h within the range of sites P and Q respectively.

[0048] Step S2: Synchronously collect the multi-frame point set of the roadside lidar sites in Step S1, merge the same sites respectively, and obtain the main roadside target domain point set P, the relay source domain point set Q, and the No. 1 roadside source domain point set M.

[0049] The roadside lidar point sets at points P, M, and Q are collected. Based on the onboard ROS system, 10 frames of point clouds at the three locations are collected simultaneously and independently within 1 second, forming sets P, M, Q∈R. n×3 .

[0050] Step S3: Preprocess the point cloud sets separately.

[0051] Preprocessing operations are performed on point sets P, M, and Q respectively, including at least removing duplicate point clouds, filtering for device noise, irrelevant point clouds, and ground points. The processing steps for each point set are as follows:

[0052] Step S31: Use Gaussian filtering to remove equipment noise.

[0053] Using a point set P(x,y,z) as input, G as a Gaussian function, a window radius of r, and a standard deviation of σ for the Gaussian function, the filtering formula is as follows:

[0054]

[0055] Where P′ is the filtered point cloud set, w is the total weight, N is the number of points within the window radius r, and p is the coordinate of the current point. i These are the coordinates of each point within the window.

[0056] Step S32: Use statistical filtering to remove irrelevant point clouds.

[0057] Use the KD-Tree to find all points N within the filter radius r. i ={p j |||p i -p j ||≤r}, and calculate its mean and standard deviation:

[0058]

[0059] By combining point cloud curvature and point cloud density values, a dynamic threshold T is set to determine p. iWhether it is located inside the point cloud: When the standard deviation is less than the threshold, the curvature and normal vector around the point are relatively average and it is retained; otherwise, the average distance to the adjacent points exceeds the defined interval, the point is considered an outlier and deleted.

[0060] Step S33: Plane fitting, remove ground points from the P, M, and Q point sets.

[0061] Let each point in the point set P be represented as (x i ,y i ,z i The ground equation is expressed as z = ax + by + c. To solve for the optimal fitted ground parameters, we need to minimize the distance D from all points to the fitted plane, i.e.:

[0062]

[0063] Taking the partial derivatives of a, b, and c with respect to D, and setting the partial derivatives to 0, we obtain the optimal solution system of equations:

[0064]

[0065] Simplify to matrix form:

[0066]

[0067] To avoid the inefficiency of calculating the entire point set, RANSAC is used to randomly sample m points from the point set P that are not on the same straight line. The minimum point set is expressed as:

[0068] S←RandomSample(P,m)

[0069] The least squares method is used to fit the ground model. A distance threshold D' is set, and the number of local points of the current ground model in the point cloud P is calculated. The values ​​of a, b, and c obtained under the maximum number of local points are retained until the iteration number K is reached.

[0070] Step S34: Coordinate correction.

[0071] Based on the plane fitting results, with the ground as the z-axis reference, the point set P is rotated according to the rotation vector (-b,-a,0). The ground points are then substituted into the plane equation to obtain the ground height h. The point cloud is then translated according to the vector (0,0,h) so that the z-coordinate of the ground points is 0, which facilitates the subsequent fitting of the spatial representation equation of the active features.

[0072] Step S4: Select the region to be registered.

[0073] To avoid excessive irrelevant and homogeneous point clouds affecting registration in tunnel scenes, point cloud regions containing actively deployed geometric feature identifiers are manually selected, such as... Figure 2As shown, the selection criteria are approximately 2 meters before and after the active feature in the x-direction under the lidar coordinate system, with no restrictions in the y and z directions.

[0074] Step S5: NDT coarse matching of initial pose.

[0075] The NDT algorithm is used for coarse registration of the selected region's source point cloud Q and target point cloud P. The detailed algorithm flow is as follows: Figure 3 As shown, the main steps include grid division, statistical value calculation, normal distribution function construction and joint density prediction. The iterative operation finally solves for the rotation and translation parameters. The detailed sub-steps are as follows.

[0076] Step S51: Divide the source point cloud into 3D raster cells and calculate the mean and covariance of the raster points:

[0077]

[0078] Where N v q represents the number of electric charges in the grid. i Let i be the i-th point in the 3D grid;

[0079] Step S52: Based on the initial pose transformation method T, for point q i Perform rotation and translation transformations:

[0080] q i =T(τ,q) i )=Rq i +t

[0081] Where τ is the transformation parameter, R is the rotation matrix, t is the translation vector, and the new coordinate point after the initial transformation is q. i ';

[0082] Step S53: Integrate the rotated and translated points with the target point in the same coordinate system, and calculate the probability that the new point cloud obtained according to the normal distribution falls into the reference point cloud, that is, the overlap function f between the source point cloud and the reference point cloud after the T-transformation:

[0083]

[0084] Then, by summing the probability densities calculated for all the graticules, the NDT registration score is obtained:

[0085]

[0086] Step S54: Optimize the objective function using the Newton optimization algorithm, obtain the transformation parameter τ with the highest registration score, and perform coarse registration of the point cloud based on the transformation parameter.

[0087] Step S6: Fit active key points.

[0088] Fit the center of mass of the rigid body to the center points of the top and bottom. The selected active features are three cylinders with radius r and heights h, h, and 2h, respectively.

[0089] Based on the ground fitting and correction results, the equation of the side surface of a cylinder with radius r and height h is expressed as:

[0090]

[0091] Where (A, B) are the coordinates of the center of the circle on the ground (xoy plane). According to the correction result of step S35, the ground point cloud height is 0. The active feature point set P that falls on the cylinder is P_active_feature_point_set. C {x i ,y i ,z i Let the distance from the point to the center of the circle be d. i Based on the least squares method to limit the error, the objective function is defined as follows:

[0092]

[0093] Setting the partial derivatives of the objective function with respect to A and B to 0, we find the extreme values ​​of the objective function and solve for A and B. Then, the centroid of this active feature marker and the three-dimensional spatial point at the center of the top and bottom edges are... (A,B,0) and (A,B,h). Similarly, the spatial coordinates of the centroid and the centers of the upper and lower bases of the other two geometric features can be obtained.

[0094] Step S7: Perform fine registration of the point cloud using point-to-point ICP computation based on singular value decomposition.

[0095] Step S71: Based on the three fitted active geometric feature points obtained in step S6, the target point cloud P can be obtained. * {p1,p2,...,p n}, Q * {q1,q2,...,q n}, where n = 9. Suppose that the transformation matrix T*(R*, t*) has the ability to map the point cloud P* to Q*. Calculate the centroid and covariance matrices of P* and Q* as follows:

[0096]

[0097] Step S72: Perform SVD singular value decomposition on the covariance, cov pq =U∑V T Where U and V are the left and right singular matrices, respectively, and ∑ is a diagonal matrix containing singular values, the rotation matrix R* can be obtained as follows:

[0098]

[0099] Where det is the determinant calculation, and the translation vector t* is calculated as follows:

[0100]

[0101] The final solution is the transformation matrix:

[0102]

[0103] Step S8: Iterate over the remaining LiDAR points to complete pose matching.

[0104] Using the relay lidar as the target point cloud and the lidar to be registered (No. 1) as the source point cloud, repeat steps S4 to S7 to obtain the rotation and translation matrix T1 between the relay lidar point cloud and the lidar to be registered (No. 1). Then, based on the transformation matrix T1 between the main lidar and the relay lidar... * The pose transformation matrix T2 of the point cloud of the main lidar and lidar No. 1 can then be obtained:

[0105] T2 = T* × T1

[0106] By removing the relay lidar and redeploying active geometric features and relay lidars at lidar points 2, 3, ..., n, the relative pose relationships of all roadside lidar points can be calculated sequentially.

[0107] In summary, the method of this invention addresses the problem of inaccurate registration caused by sparse overlap of long-distance roadside lidar. It utilizes mobile lidar as an auxiliary relay to enhance point density distribution. By arranging multiple geometric feature markers, it fits the three-dimensional spatial representation of the geometric object, thereby solving for the spatial coordinates of the rigid body's centroid and the centers of its top and bottom edges, forming point sets in the source and target domains. This effectively avoids the distribution differences of point clouds with different densities on the same active feature. By combining NDT coarse matching with ICP fine matching based on active features, this method not only effectively improves the convergence time but also obtains high-precision point-to-point matching rotation and translation matrix parameters. It effectively overcomes the problem of difficult registration of long-distance roadside lidar within the perception range, and better serves accurate perception and assisted positioning in tunnel scenarios such as satellite rejection and point cloud homogenization.

[0108] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A roadside multi-liquid radar registration method based on active features and mobile relay, characterized in that, Includes the following steps: S1, Deploy roadside locations: The roadside locations include lidar locations and geometric active features, wherein the lidar locations include a main lidar P, a mobile relay lidar Q, and a lidar M to be registered. S2, Continuous sampling point set: Collect the roadside lidar point sets at the main lidar P, mobile relay lidar Q and lidar M to be registered, merge the same points respectively, and obtain the main roadside target domain point set P, the relay source domain point set Q and the roadside source domain point set M. S3, Point Cloud Set Preprocessing: Perform preprocessing operations on the point sets P, M, Q obtained in step S2. The preprocessing includes at least removing device noise, extracting irrelevant point clouds, plane fitting, and coordinate correction. S4, Select active feature region: Manually select a point cloud region containing actively deployed geometric feature identifiers. The selection criterion is approximately 2 meters before and after the active feature in the x direction under the lidar coordinate system, with no restrictions in the y and z directions. S5, NDT coarse matching initial pose: The NDT algorithm is used to perform coarse registration on the region selected in step S4 to determine the key point region for precise registration; S6, Fitting Active Keypoints: Based on the three common geometric features deployed under the two point clouds to be registered, solve the spatial representation equations of each geometric feature under the source point cloud and the target point cloud respectively. Taking a cylinder with radius r and height h as an example, its spatial representation equation is: Where (A, B) are the coordinates of the center of the circle on the ground (xoy plane); according to the coordinate correction result in step S3, the ground point cloud height is 0. Based on the least squares method to limit the error, the three-dimensional spatial points of the centroid and the upper and lower bottom centers of the active feature marker are calculated as follows: Similarly, find the spatial coordinates of the centroid and the center of the upper and lower bases of the other two geometric features (A,B,0) and (A,B,h). S7: ICP fine matching based on SVD decomposition: Point cloud fine registration based on point-to-point calculation of singular value decomposition; S8: Pose transfer and positioning broadcast: Repeat steps S4-S7, iterate through all LiDAR points, complete pose matching, and calculate the relative pose relationship of all roadside LiDAR points in sequence through the rotation and translation relationship between two adjacent LiDARs.

2. The roadside multi-liquid radar registration method based on active features and mobile relay as described in claim 1, characterized in that: The point cloud set preprocessing in step S3 specifically includes: S31: Using a point set as input, Gaussian filtering is used to remove device noise; S32: Use KD-Tree to find all point sets within a filtering radius r, calculate the standard deviation of point cloud curvature and density based on statistical filtering, and set a threshold to remove irrelevant points; S33: Fit the ground equation using the least squares method to obtain the parameters of the ground equation z=ax+by+c, and remove the ground points from each lidar point set; S34: Fit the height of the point cloud on the z-axis using the ground equation, and perform coordinate transformation on each point set based on the rotation vector obtained in step S33 and the translation vector obtained in this step to perform coordinate correction.

3. The roadside multi-liquid radar registration method based on active features and mobile relay as described in claim 2, characterized in that: In the Gaussian filtering step S31, the point set P(x,y,z) is used as input, G is a Gaussian function, the window radius is r, and the standard deviation of the Gaussian function is σ. The filtering formula is as follows: Where P′ is the filtered point cloud set, w is the total weight, N is the number of points within the window radius r, and p is the coordinate of the current point. i These are the coordinates of each point within the window.

4. The roadside multi-liquid radar registration method based on active features and mobile relay as described in claim 1, characterized in that: The specific steps of step S5, NDT coarse matching initial pose, include the following: S51: Divide the source point cloud into three-dimensional grids and calculate the mean and covariance of each grid point; S52: Rotate and translate the source point cloud according to the initialized pose transformation parameters; S53: Integrate the rotated and translated points with the target point in the same coordinate system, and calculate the probability that the new point cloud obtained according to the normal distribution falls within the reference point cloud. The overlap function f between the source point cloud and the reference point cloud after the T-transform is: Where q i The source point cloud is μ, the centroid of the point cloud is cov, and the point cloud covariance is covariance. Then, the NDT registration score is obtained by summing the probability densities calculated from all the grids. S54: The objective function is optimized using the Newton-Raphson optimization algorithm to obtain the transformation parameter τ with the highest registration score.

5. The roadside multi-liquid radar registration method based on active features and mobile relay as described in claim 1, characterized in that: Step S7, point cloud fine registration based on point-to-point calculation using singular value decomposition, specifically includes: S71: Based on the three fitted active geometric feature points obtained in step S6, the target point cloud P can be obtained. * {p1,p2,...,p n }, Q * {q1,q2,...,q n Where n = 9, the centroid and covariance are calculated based on the point set; S72: Perform SVD singular value decomposition on the covariance to obtain the rotation matrix R*, thereby deriving the translation vector t* and obtaining the transformation matrix T. * .

Images