A Single Radar Calibration and Rigid-Body-Based Multi-Radar Registration Method

Abstract

The present invention belongs to the technical field of radar calibration and registration, and specifically relates to a single radar calibration and a multi-radar registration method based on rigid bodies. The multi-radar registration includes the following steps: P1 Install multiple lidars at different positions at the application scenario site, and select a ground recognition rigid body between every two adjacent lidars; P2 Calibrate each lidar respectively; P3 Point cloud cube interception; P4 Generate the point cloud rotation and translation matrices of adjacent lidars in sequence; P5 Multi-radar point cloud fusion; Compared with the prior art, (1) Both the single radar calibration and the multi-radar registration method based on rigid bodies of the present invention can be realized by programming, eliminating the relatively cumbersome manual processing process required for multi-radar point cloud registration, and improving the efficiency and accuracy; (2) In the multi-radar registration method based on rigid bodies of the present invention, by selecting a ground recognition rigid body between every two adjacent lidars, it is convenient to generate the point cloud rotation and translation matrices of adjacent lidars in sequence.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar calibration and registration, and particularly relates to a single radar calibration and a multi-radar registration method based on a rigid body. Background Art

[0002] The lidar point cloud is default with the radar centroid as the coordinate origin (radar original coordinate system). In actual applications, it is generally required to use the center of the application scenario on the ground as the coordinate origin (custom coordinate system). This requires calibration of the lidar to complete the coordinate transformation of the point cloud. This usually requires multiple calibration attempts, and the calibration process lacks an engineering method.

[0003] The effective coverage diameter of the lidar point cloud that meets the resolution requirements is usually relatively limited. For example: for a 32-line lidar, if the installation height is 1.5 m from the scanned object, its maximum coverage diameter can reach 74 m. However, if the point cloud resolution is required not to exceed 15 cm, the effective coverage diameter is only 3 m.

[0004] This often fails to meet the needs of the application scenario (such as the application scenario of automatic loading of bulk materials such as coal). For example, in the above application scenario, it is necessary to determine the length, width, and height of the loading vehicle, locate the edge of the car body, determine the moving distance of the loading vehicle, and detect the feeding height, etc. At this time, it is necessary to install multiple lidars at appropriate intervals, and fuse the radar point clouds into a single integrated point cloud data, which can cover a larger range and further improve the point cloud resolution. Precise multi-radar point cloud registration usually also requires a relatively cumbersome manual processing process. Summary of the Invention

[0005] In view of the problem that precise multi-radar point cloud registration usually requires a relatively cumbersome manual processing process, the present invention provides a single radar calibration and a multi-radar registration method based on a rigid body.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a single radar calibration method, where the single radar calibration is a process of adjusting the original three-dimensional point cloud coordinate system of a single lidar to a target pose. In the target pose, the three-dimensional point cloud coordinate system of the lidar uses the ground plane as the XY plane of the three-dimensional point cloud coordinate system, and uses the center of the application scenario of the lidar on the ground plane as the coordinate origin of the three-dimensional point cloud coordinate system;

[0007] The single radar calibration method includes the following steps:

[0008] S1 Obtain the original three-dimensional point cloud output by the lidar and perform noise reduction processing on the original three-dimensional point cloud;

[0009] S2 Extract the ground plane from the original three-dimensional point cloud output by the lidar;

[0010] S3 rotates and translates the three-dimensional point cloud output by the lidar to transform the three-dimensional point cloud to the target pose and saves it, completing the single lidar calibration work.

[0011] Preferably, in S1, the lidar is installed in the application scenario, the original three-dimensional point cloud output by the lidar is recorded in the non-production state, and the standard voxel grid filter in the open-source point cloud processing library PCL is used to filter the miscellaneous points in the original three-dimensional point cloud, completing the noise reduction of the original three-dimensional point cloud.

[0012] Preferably, the extraction of the ground plane from the original three-dimensional point cloud in S2 is carried out in four steps:

[0013] In the coordinate system of the lidar's original three-dimensional point cloud in S21, starting from the coordinate origin, the angles are equally divided in the vertical and horizontal directions, and the equally divided lines are rays. A family of rays in the coordinate system of the original three-dimensional point cloud is formed; the equally divided angle is a, and the number of rays in the family of rays is N:

[0014] The square brackets indicate taking the integer.

[0015] The zenith angle θ of the polar coordinates corresponding to each equally divided line i and the azimuth angle are respectively:

[0016] θ i = a×i, where i is an integer from 0 to ;

[0017] j is an integer from 0 to ;

[0018] In S22, the pixel coordinates in the coordinate system of the original three-dimensional point cloud are converted to polar coordinates, and the conversion formula is:

[0019]

[0020] In the conversion formula, r is the distance from the pixel point to the coordinate origin in the coordinate system of the original three-dimensional point cloud; θ is the zenith angle of the pixel point in the polar coordinates; is the azimuth angle of the pixel point in the polar coordinates;

[0021] In S23, k rays are selected from the family of rays, and a pixel point is selected on each selected ray; the pixel point farthest from the coordinate origin in the open area enclosed by four adjacent rays is selected as the base point, and the number of the finally taken base points is denoted as K;

[0022] S24 uses the Random Sample Consensus algorithm RANSAC in the open-source point cloud processing library PCL to extract planes for each of the K base points respectively. The plane with the largest area among the extracted planes is the ground plane.

[0023] Preferably, in step S3, the center point of the extracted ground plane is used as the coordinate origin of the custom coordinate system. Then, the normal vector of the ground plane is calculated, and the matrix multiplication with column vectors in the open-source point cloud processing library PCL is used to calculate the transformation of the three-dimensional point cloud to the target pose.

[0024] A multi-radar registration method based on a rigid body based on the single radar calibration method includes the following steps:

[0025] P1 Install multiple lidars at different positions at the application scenario site of the lidar. Select a ground recognition rigid body between every two adjacent lidars; if there is no ground recognition rigid body available for selection, place a traffic cone between every two adjacent lidars as the ground recognition rigid body.

[0026] P2 Use the single radar calibration method to complete the calibration of each lidar respectively.

[0027] P3 Intercept a point cloud cube in the three-dimensional point cloud coordinate system of the calibrated lidar. The intercepted point cloud cube includes the ground plane and the ground recognition rigid body for recognition.

[0028] P4 Generate the point cloud rotation and translation matrices of adjacent lidars in sequence; the same ground recognition rigid body in the application scenario is included in the point clouds of adjacent lidars.

[0029] P5 Multi-radar point cloud fusion. Designate one lidar as the reference lidar. The point cloud coordinates of the lidar adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix one to be transformed into the point cloud coordinates of the reference lidar; for the point cloud transformation of the lidar not adjacent to the reference lidar, first transform the point cloud coordinates of the lidar not adjacent to the reference lidar into the point cloud coordinates of the lidar adjacent to the reference lidar, and then multiply the point cloud coordinates of the lidar adjacent to the reference lidar by the corresponding point cloud rotation and translation matrix one to be transformed into the point cloud coordinates of the reference lidar, finally completing the unification of the multi-radar coordinate systems.

[0030] The point cloud coordinates of the lidar not adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix two to be transformed into the point cloud coordinates of the lidar adjacent to the reference lidar.

[0031] Preferably, in P4, the point clouds of two adjacent lidars are fed into a neural network model. The backbone network of the neural network model is a three-dimensional convolutional neural network, and the three-dimensional convolutional neural network includes twenty-one three-dimensional convolutional layers (conv3d); the neural network model extracts regional features from the input point clouds of the two lidars respectively; the twenty-one three-dimensional convolutional layers are divided into five groups, denoted as group C1 to group C5 respectively, and convolutions are performed within the corresponding groups respectively. The output of the previous group is the input of the next group, and each group is followed by a three-dimensional pooling (pool3d) operation after multiple three-dimensional convolution operations;

[0032] C1: conv3d11(3, 16), conv3d12(3, 16), conv3d13(3, 16), pool3d1(2, 2), bn3d1

[0033] C2: conv3d21(3, 64), conv3d22(3, 64), conv3d23(3, 64), pool3d2(2, 2), bn3d2

[0034] C3: conv3d31(3, 128), conv3d32(3, 128), conv3d33(3, 128), conv3d34(3, 128), conv3d35(3, 128), pool3d3(2, 2)

[0035] C4: conv3d41(3, 256), conv3d42(3, 256), conv3d43(3, 256), conv3d44(3, 256), conv3d45(3, 256), pool3d4(2, 2)

[0036] C5: conv3d51(3, 512), conv3d52(3, 512), conv3d53(3, 512), conv3d54(3, 512), conv3d55(3, 512), pool3d5(2, 2)

[0037] Among them, the first parameter in the three-dimensional convolutional layer represents the convolution kernel size (e.g., 3 means the convolution kernel size is 3×3×3), the second parameter represents the output dimension, the input dimension is the output dimension of its previous operation, and the input dimension of the first three-dimensional convolutional layer is 1;

[0038] After the 3D pooling (pool3d) operation of groups C1 and C2, a batch normalization (bn3d, where bn is the abbreviation of BatchNormalization) operation is followed to perform mean normalization on the feature data. After the batch normalization (bn3d) operation, the normalized mean is output to the next group.

[0039] The process of performing mean normalization on the feature data in groups C1 and C2 is as follows:

[0040]

[0041]

[0042] Among them, x i is the input data of each convolutional layer, y is the output data of each convolutional layer, μ is the 3D mean, σ is the 3D standard deviation, and m is the number of input data of each convolutional layer.

[0043] After the batch normalization (bn3d) operation, the normalized mean is close to zero, eliminating the similarity influence of flat regions such as the ground and retaining the similarity influence highlighting the rigid body region for ground recognition.

[0044] Then, calculate the similarity of the C5 group feature outputs of the two point clouds:

[0045]

[0046] Among them, F is the regional feature vector, n is the feature dimension, and the value range of the similarity S is from 0 to 2. The closer the value is to 0, the closer the features of the two regions are, and record the numbers of the regions (i, j) with the smallest S value in each round of training.

[0047] Extract the region with the smallest similarity S according to the recorded (i, j) numbers, and inversely map the corresponding point cloud regions of the rigid body for ground recognition in the point clouds of the two lidars through deconvolution. The 3D coordinates corresponding to the upper left corner of the corresponding point cloud region are the 3D coordinates of the upper left corner of the same rigid body for ground recognition in the point clouds of the two adjacent lidars. Since the two adjacent lidars have completed calibration, only translation is required in the z direction for the two coordinate systems, and thus the point cloud rotation and translation matrix can be calculated based on a set of coordinate pairs.

[0048] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0049] (1) Using the single lidar calibration method of the present invention, a ray family is formed by equally dividing the angles of the original three-dimensional point cloud coordinate system in the vertical and horizontal directions; the pixel coordinates in the original three-dimensional point cloud coordinate system are subjected to polar coordinate transformation; K base points are selected from the ray family, and the random sample consensus algorithm RANSAC in the open-source point cloud processing library PCL is used for each of the K base points to extract a plane, and the plane with the largest area among the extracted planes is the ground plane;

[0050] (2) Rotate and translate the three-dimensional point cloud output by the lidar to transform the three-dimensional point cloud to the target pose and save it to complete the single lidar calibration work; the center of the extracted ground plane is used as the coordinate origin of the custom coordinate system in the target pose;

[0051] (3) In the multi-lidar registration method based on rigid bodies of the present invention, multiple lidars are installed at different positions at the application scenario site, and a ground recognition rigid body is selected between every two adjacent lidars; the point cloud rotation and translation matrices of adjacent lidars are generated in sequence, and the same ground recognition rigid body in the application scenario is included in the point clouds of adjacent lidars; in the multi-lidar point cloud fusion, one lidar is designated as the reference lidar, and the point cloud of the lidar adjacent to the reference lidar is multiplied by the corresponding point cloud rotation and translation matrix to be transformed to the pose of the reference lidar; the transformation of the point cloud of the lidar not adjacent to the reference lidar is achieved by sequentially multiplying the point cloud of the lidar by the point cloud rotation and translation matrices of adjacent lidars, and finally the unification of the multi-lidar coordinate systems is completed;

[0052] (4) In the process of sequentially generating the point cloud rotation and translation matrices of adjacent lidars, the point clouds of two adjacent lidars are sent into the neural network model to extract the regional features of the ground recognition rigid body, and then the similarity comparison of the regional features in the point clouds of two adjacent lidars is completed;

[0053] (5) The single lidar calibration and the multi-lidar registration method based on rigid bodies of the present invention can both be realized by programming, eliminating the relatively cumbersome manual processing process required for multi-lidar point cloud registration, and improving the efficiency and accuracy. Description of the Drawings Detailed Embodiments

[0054] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described below with reference to the embodiments.

[0055] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0056] Embodiment 1

[0057] The present invention will be further described below. A multi-radar calibration and registration method based on a rigid body. The single-radar calibration is a process of adjusting the original three-dimensional point cloud coordinate system of a single lidar to a target pose. In the target pose, the XY plane of the three-dimensional point cloud coordinate system of the lidar is the ground plane, and the coordinate origin of the three-dimensional point cloud coordinate system is the application scenario center of the lidar on the ground plane.

[0058] The single-radar calibration method includes the following steps:

[0059] S1: Obtain the original three-dimensional point cloud output by the lidar and perform noise reduction processing on the original three-dimensional point cloud;

[0060] S2: Extract the ground plane from the original three-dimensional point cloud output by the lidar;

[0061] S3: Rotate and translate the three-dimensional point cloud output by the lidar to transform the three-dimensional point cloud to the target pose, and save it to complete the single-radar calibration work.

[0062] In S1, install the lidar in the application scenario, record the original three-dimensional point cloud output by the lidar in the non-production state, and use the standard voxel grid filter in the open-source point cloud processing library PCL to filter out the noise points in the original three-dimensional point cloud to complete the noise reduction of the original three-dimensional point cloud.

[0063] The extraction of the ground plane from the original three-dimensional point cloud in S2 is carried out in four steps:

[0064] S21: In the lidar original three-dimensional point cloud coordinate system, starting from the coordinate origin, perform angular equal division along the vertical and horizontal directions. The equal division lines are rays, and multiple equal division lines form a ray family in the lidar original three-dimensional point cloud coordinate system; the equal division angle is a, and the number of rays in the ray family is N:

[0065] The square brackets indicate taking the integer

[0066] The zenith angle θ of the polar coordinates corresponding to each equal division line i and the azimuth angle are respectively:

[0067] θ i = a × i,

[0068]

[0069] S22: Perform polar coordinate conversion on the pixel coordinates in the original three-dimensional point cloud coordinate system. The conversion formula is:

[0070]

[0071] In the conversion formula, r is the distance from the pixel point to the coordinate origin in the original three-dimensional point cloud coordinate system; θ is the zenith angle of the pixel point in polar coordinates; is the azimuth angle of the pixel point in polar coordinates; x, y, and z are the x coordinate, y coordinate, and z coordinate of the pixel point.

[0072] S23 Select k rays from the ray family, and select a pixel point on each selected ray; select the pixel point farthest from the coordinate origin within the open area enclosed by four adjacent rays as the base point, and the number of base points finally taken out is denoted as K;

[0073] r k =max (r ij )

[0074] θ k ∈[θ i ,θ i+1

[0075]

[0076] where k <= N, and the above algorithm can quickly obtain the pixel point farthest from the coordinate origin from the open area enclosed by four adjacent rays (if there is no pixel point in this area, it is vacant).

[0077] S24 Use the random sample consensus algorithm RANSAC in the open source point cloud processing library PCL to extract planes for each of the K base points, and the plane with the largest area among the extracted planes is the ground plane.

[0078] In step S3, take the center point of the extracted ground plane as the coordinate origin of the custom coordinate system, then calculate the normal vector of the ground plane, and then use the calculation method of matrix multiplying column vector in the open source point cloud processing library PCL to calculate the transformation of the three-dimensional point cloud to the target pose.

[0079] A multi-radar registration method based on a rigid body based on the above single-radar calibration method includes the following steps:

[0080] P1 Install multiple lidars at different positions at the application scenario site of the lidar, and select a ground recognition rigid body between every two adjacent lidars; if there is no available ground recognition rigid body, place a traffic cone between every two adjacent lidars as the ground recognition rigid body;

[0081] P2 Use the single-radar calibration method to complete the calibration of each lidar respectively;

[0082] P3 Intercept a point cloud cube in the three-dimensional point cloud coordinate system of the calibrated lidar, and the intercepted point cloud cube includes the ground plane and the ground recognition rigid body for identification;

[0083] ​P4 generates the point cloud rotation and translation matrices for adjacent lidars in sequence; the point clouds of adjacent lidars contain the same ground recognition rigid body in the application scenario.

[0084] P5 performs multi-lidar point cloud fusion. Designate one lidar as the reference lidar. The point cloud coordinates of the lidars adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix I generated in P4 and transformed to the pose of the reference lidar. For the point cloud transformation of the lidars not adjacent to the reference lidar, first transform the point cloud coordinates of the lidars not adjacent to the reference lidar to the point cloud coordinates of the lidars adjacent to the reference lidar, and then multiply the point cloud coordinates of the lidars adjacent to the reference lidar by the corresponding point cloud rotation and translation matrix I (generated in P4) and transform to the pose of the reference lidar, finally completing the unification of the multi-lidar coordinate systems.

[0085] The point cloud coordinates of the lidars not adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix II (generated in P4) and transformed to the point cloud coordinates of the lidars adjacent to the reference lidar.

[0086] In P4, the point clouds of two adjacent lidars are fed into the neural network model. The backbone network of the neural network model is a three-dimensional convolutional neural network, and the three-dimensional convolutional neural network contains twenty-one three-dimensional convolutional layers (conv3d). The neural network model extracts regional features from the input point clouds of the two lidars respectively. The twenty-one three-dimensional convolutional layers are divided into five groups, denoted as group C1 to group C5 respectively, and perform convolutions within the corresponding groups. The output of the previous group is the input of the next group, and each group is followed by a three-dimensional pooling pool3d operation after multiple three-dimensional convolution operations.

[0087] C1: conv3d11(3,16), conv3d12(3,16), conv3d13(3,16), pool3d1(2,2), bn3d1

[0088] C2: conv3d21(3,64), conv3d22(3,64), conv3d23(3,64), pool3d2(2,2), bn3d2

[0089] C3: conv3d31(3,128), conv3d32(3,128), conv3d33(3,128), conv3d34(3,128), conv3d35(3,128), pool3d3(2,2)

[0090] C4: conv3d41(3,256), conv3d42(3,256), conv3d43(3,256), conv3d44(3,256), conv3d45(3,256), pool3d4(2,2)

[0091] C5: conv3d51(3, 512), conv3d52(3, 512), conv3d53(3, 512), conv3d54(3, 512), conv3d55(3, 512), pool3d5(2, 2)

[0092] Among them, conv is the abbreviation of Convolution; the first parameter in the three-dimensional convolution layer represents the convolution kernel size (e.g., 3 means the convolution kernel size is 3×3×3), the second parameter represents the output dimension, and the input dimension is the output dimension of the previous operation. The input dimension of the first three-dimensional convolution layer is 1.

[0093] The bn3d (bn is the abbreviation of Batch Normalization) operation is added to both groups of C1 and C2 to perform mean normalization on the feature data, as follows:

[0094]

[0095]

[0096] Among them, x i is the input data of each convolution layer, y is the output data of each convolution layer, μ is the three-dimensional mean, σ is the three-dimensional standard deviation, and m is the number of input data of each convolution layer;

[0097] After the bn3d (bn is the abbreviation of Batch Normalization, which means batch normalization) operation, the similarity influence of plane regions such as the ground is eliminated (the mean value approaches zero after normalization), and the similarity influence of the rigid body region highlighting the ground recognition is retained;

[0098] Then, the similarity of the C5 group feature outputs of the two point clouds is calculated:

[0099]

[0100] Among them, F is the regional feature vector, n is the feature dimension, the value range of the similarity S is [0, 2], the closer the value is to 0, the closer the features of the two regions are, and the numbers of the regions (i, j) with the minimum S value in each round of training are recorded;

[0101] Extract the region with the smallest similarity S according to the recorded (i, j) numbers, and use deconvolution to inversely map the point cloud regions corresponding to the ground recognition rigid body in the point clouds of the two lidars respectively. The three-dimensional coordinates corresponding to the upper left corner of the corresponding point cloud region are the three-dimensional coordinates of the upper left corner of the same ground recognition rigid body in the point clouds of the two adjacent lidars. Since the two adjacent lidars have been calibrated, only translation is required in the z direction for the two coordinate systems. In this way, the point cloud rotation and translation matrix can be calculated based on a set of coordinate pairs.

[0102] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification or equivalent change made to the above embodiments based on the technical essence of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A single radar calibration method, characterized in that, The single lidar calibration is a process of adjusting the original three-dimensional point cloud coordinate system of a single lidar to the target pose. In the target pose, the XY plane of the three-dimensional point cloud coordinate system of the lidar is the ground plane, and the coordinate origin of the three-dimensional point cloud coordinate system is the center of the application scenario of the lidar on the ground plane; The single lidar calibration method includes the following steps: S1 Obtain the original three-dimensional point cloud output by the lidar and perform noise reduction processing on the original three-dimensional point cloud; S2 Extract the ground plane from the original three-dimensional point cloud output by the lidar. The extraction of the ground plane from the original three-dimensional point cloud in S2 is carried out in four steps: S21 In the lidar original three-dimensional point cloud coordinate system, starting from the coordinate origin, perform angular equal division along the vertical and horizontal directions. The equal division lines are rays, and multiple equal division lines form a ray family in the lidar original three-dimensional point cloud coordinate system; The equal division angle is a, and the number of rays in the ray family is N: , The square brackets indicate rounding; The zenith angle of the polar coordinates corresponding to each of the bisecting lines and the azimuth angle are respectively: S22 Perform polar coordinate transformation on the pixel coordinates in the original three-dimensional point cloud coordinate system. The transformation formula is: ) In the conversion formula, r is the distance from the pixel point to the origin of coordinates in the original three-dimensional point cloud coordinate system; is the zenith angle of the pixel point in polar coordinates; is the azimuth angle of the pixel point in polar coordinates; S23 Select k rays from the ray family, and select a pixel point on each selected ray; Select the pixel point farthest from the coordinate origin in the open area enclosed by four adjacent rays as the base point. The number of the finally taken base points is denoted as K; S24 Respectively use the random sample consensus algorithm RANSAC in the open source point cloud processing library PCL to extract the plane for each of the K base points. The plane with the largest area among the extracted planes is the ground plane; S3 Rotate and translate the three-dimensional point cloud output by the lidar to transform the three-dimensional point cloud to the target pose and save it to complete the single lidar calibration work.

2. The single radar calibration method according to claim 1, wherein In S1, install the lidar in the application scenario, record the original three-dimensional point cloud output by the lidar in the non-production state, and use the standard voxel grid filter in the open source point cloud processing library PCL to filter out the miscellaneous points in the original three-dimensional point cloud to complete the noise reduction of the original three-dimensional point cloud.

3. The single-radar calibration method according to claim 2, wherein In step S3, use the center point of the extracted ground plane as the coordinate origin of the custom coordinate system, then calculate the normal vector of the ground plane, and then use the calculation method of matrix multiplying column vectors in the open source point cloud processing library PCL to calculate the transformation of the three-dimensional point cloud to the target pose.

4. A multi-radar registration method based on rigid bodies, characterized in that, It includes the following steps: P1 Install multiple lidars at different positions on the site of the lidar application scenario, and select a ground recognition rigid body between every two adjacent lidars; If there is no available ground recognition rigid body, place a traffic cone between every two adjacent lidars as the ground recognition rigid body; P2 Use the single lidar calibration method described in claim 3 to complete the calibration of each lidar respectively; P3 Intercept a point cloud cube in the three-dimensional point cloud coordinate system of the calibrated lidar. The intercepted point cloud cube includes the ground plane and the ground recognition rigid body for recognition; P4 Generate the point cloud rotation and translation matrix of adjacent lidars in sequence; The same ground recognition rigid body in the application scenario is included in the point clouds of adjacent lidars; P5 multi-radar point cloud fusion. Designate a lidar as the reference lidar. The point cloud coordinates of the lidars adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix one to be transformed into the point cloud coordinates of the reference lidar; for the point cloud transformation of the lidars not adjacent to the reference lidar, first transform the point cloud coordinates of the lidars not adjacent to the reference lidar into the point cloud coordinates of the lidars adjacent to the reference lidar, and then multiply by the point cloud rotation and translation matrix one to be transformed into the point cloud coordinates of the reference lidar, finally completing the unification of the multi-radar coordinate systems; The point cloud coordinates of the lidars not adjacent to the reference lidar are multiplied by the corresponding point cloud rotation and translation matrix two to be transformed into the point cloud coordinates of the lidars adjacent to the reference lidar.

5. The multi-radar registration method based on a rigid body according to claim 4, wherein In P4, the point clouds of two adjacent lidars are fed into the neural network model. The backbone network of the neural network model is a three-dimensional convolutional neural network, and the three-dimensional convolutional neural network contains twenty-one three-dimensional convolutional layers; the neural network model extracts regional features from the input point clouds of the two lidars respectively; the twenty-one three-dimensional convolutional layers are divided into five groups, denoted as group C1 to group C5 respectively, and convolutions are performed within the corresponding groups. The output of the previous group is the input of the next group, and each group is followed by a three-dimensional pooling pool3d operation after multiple three-dimensional convolution operations; Both the three-dimensional pooling pool3d operations of group C1 and group C2 are followed by a batch normalization bn3d operation to perform mean normalization on the feature data; after the batch normalization bn3d operation, the normalized mean is output to the next group; Then calculate the similarity of the C5 group feature outputs of the two point clouds: Where F is the regional feature vector, n is the feature dimension, the similarity S takes values between 0 and 2, and the closer the value is to 0, the closer the features of the two regions are. Record the number of the region with the smallest S value in each round of training( , ); Extract the region with the minimum similarity S according to the recorded ( , ) number, and inversely map the corresponding point cloud regions of the ground recognition rigid body in the point clouds of the two lidars through deconvolution. The three-dimensional coordinates corresponding to the upper left corner of the corresponding point cloud region are the three-dimensional coordinates of the upper left corner of the same ground recognition rigid body in the point clouds of two adjacent lidars. In this way, the point cloud rotation and translation matrix can be calculated based on a set of coordinate pairs.

Images