A point cloud registration method, apparatus and device

Abstract

The present disclosure belongs to the technical field of computer vision, and particularly relates to a point cloud registration method, device and equipment. The method comprises: collecting point cloud data of a target region from multiple perspectives, calculating the resolution of the point cloud in each point cloud data, uniformly projecting the point cloud on a horizontal plane based on the resolution to generate a BEV image; extracting key points and establishing a corresponding relationship between the key points in the source BEV image and the target BEV image generated by the source point cloud and the target point cloud respectively, and establishing a corresponding relationship between the key points of the source point cloud and the target point cloud based on the corresponding relationship between the key points of the source BEV image and the target BEV image; and obtaining an initial value of a transformation matrix of the source point cloud and the target point cloud based on the corresponding relationship between the key points of the source point cloud and the target point cloud, and iteratively optimizing the transformation matrix.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the technical field of computer vision, and particularly relates to a point cloud registration method, device and equipment. BACKGROUND

[0002] Point cloud registration is one of the core problems in 3D computer vision and robotics, which refers to aligning two point clouds into the same coordinate system through rigid transformation. Existing point cloud registration methods have made certain progress in registration of point clouds with high similarity or small-scale scenes, but point cloud registration is still a major challenge when facing large-scale scenes. Point cloud registration of large-scale scenes has multiple challenges, including high computational cost, low accuracy, noise interference, data loss, and variability of local point distribution. In addition, insufficient overlap can lead to information loss, thereby increasing the risk of registration failure. In the face of large-scale low-overlap scenes, existing point cloud registration methods either rely on hand-crafted feature descriptors or are based on learning. The method relying on hand-crafted feature descriptors usually relies on specific structural features and is sensitive to scene complexity, while the learning-based method has low computational efficiency and high resource consumption in large-scale scenes.

[0003] Terrestrial laser scanning (TLS) point cloud and airborne laser scanning (ALS) point cloud registration is a typical case of registration problems in large-scale low-overlap scenes. Since TLS and ALS point clouds have a large data size, usually consisting of millions of points, covering an area of thousands of square meters, and due to the difference in collection angle and path, the overlap between them is often small, which brings great challenges to point cloud registration. Therefore, in order to realize the effective integration of TLS and ALS data, it is urgent to develop a point cloud registration method focusing on large-scale low-overlap scenes. SUMMARY

[0004] Embodiments of the present disclosure propose a point cloud registration scheme to provide a point cloud registration method suitable for large-scale low-overlap scenes.

[0005] A first aspect of the embodiments of the present disclosure provides a point cloud registration method, comprising:

[0006] Collecting point cloud data of a target region from multiple perspectives, calculating the resolution of the point cloud in each of the point cloud data, and uniformly projecting the point cloud onto the horizontal plane based on the resolution to generate a BEV image;

[0007] Extracting key points from the source BEV image and the target BEV image generated by the source point cloud and the target point cloud respectively, and establishing a correspondence relationship between the key points, and establishing a correspondence relationship between the key points of the source point cloud and the target point cloud based on the correspondence relationship between the key points of the source BEV image and the target BEV image;

[0008] An initial value of a transformation matrix of the source point cloud and the target point cloud is obtained based on the correspondence of the key points of the source point cloud and the target point cloud, and the transformation matrix is iteratively optimized.

[0009] In some embodiments of the present disclosure, the resolution of the point cloud comprises:

[0010]

[0011] wherein RES represents the resolution of the point cloud, N represents the number of points in the point cloud, γ is a hyperparameter, x max , x min , y max , y min respectively represent the maximum and minimum values of the point cloud in the X and Y directions.

[0012] In some embodiments of the present disclosure, the uniform projection of the point cloud on the horizontal plane based on the resolution comprises:

[0013] extracting the horizontal plane in the point cloud, obtaining the angle between the horizontal plane and the XOY plane; rotating the point cloud based on the angle, so that the horizontal plane is parallel to the XOY plane;

[0014] adjusting the distance between points in the point cloud, so that the point cloud is uniformly projected on the XOY plane based on the resolution.

[0015] In some embodiments of the present disclosure, the generation of the BEV image comprises:

[0016] establishing a mapping from the point cloud coordinate system to the pixel coordinate system:

[0017]

[0018] wherein D ij represents a set of points projected to a pixel at pixel coordinate (i, j), P is a point cloud with N points, wherein the X and Y coordinates of the kth point p k are x k and y k , x min and y min are the minimum values of the X and Y directions in the point cloud coordinate system, respectively;

[0019] based on the maximum height value H ij of the points in D ij , a BEV image matrix G is generated, wherein each element G ij corresponds to the gray value of the pixel at coordinate (i, j),

[0020]

[0021]

[0022] wherein H ij is D ij the maximum height value of the midpoint, G ij is the gray value of the pixel at coordinate (i, j) of the BEV image matrix G, D ij represents the point set projected to the pixel at pixel coordinate system coordinate (i, j), z max and z min are the maximum value and the minimum value of the Z direction of the point cloud coordinate system respectively, the height value of the kth point p k is z k .

[0023] In some embodiments of the present disclosure, after the BEV image is generated, further comprising:

[0024] using two different intensity high-pass filters to enhance the high-frequency content in the image, and the filtering operation is represented as:

[0025] wherein,

[0026]

[0027] is the enhanced image matrix, w1 represents the convolution kernel for extracting edges, w2 represents the convolution kernel for sharpening details, and ★ represents the convolution operation.

[0028] In some embodiments of the present disclosure, the key points are extracted and the correspondence between the key points is established, comprising:

[0029] extracting the key points of the BEV image by using SuperPoint;

[0030] pairing the key points of different BEV images generated by different point clouds of the target region through the image key point matching framework LightGlue;

[0031] if the number of matching key points with pairing relationship is lower than a preset threshold, performing key point extraction and matching on the neighborhood of the matching key points again.

[0032] In some embodiments of the present disclosure, the correspondence between the key points of the source point cloud and the target point cloud is established based on the correspondence between the key points of the source BEV image and the target BEV image, comprising:

[0033] mapping the matching key points of the source BEV image and the target BEV image to the source point cloud and the target point cloud respectively, to obtain point clouds with the correspondence, wherein the mapping relationship is:

[0034]

[0035] RES represents the resolution of the point cloud, X, Y, Z are the coordinates of the point cloud points in the three-dimensional coordinate system, G ij Corresponding to the pixel gray value of the BEV image at coordinates (i, j), x min and y min are the minimum values of the point cloud coordinate system X and Y directions, respectively, z max and z min are the maximum and minimum values of the point cloud coordinate system Z direction, respectively.

[0036] In some embodiments of the present disclosure, the iterative optimization of the transformation matrix comprises:

[0037] Input the correspondence of the point cloud represented by the transformation matrix into the SVD process, and perform iteration to update the transformation matrix until the registration of the source point cloud and the target point cloud based on the transformation matrix reaches a preset standard;

[0038] Input the transformation matrix into the ICP algorithm to achieve further registration.

[0039] A second aspect of an embodiment of the present disclosure provides a point cloud registration device, comprising:

[0040] A projection module is configured to collect point cloud data of a target region from multiple perspectives respectively, and in each of the point cloud data, the resolution of the point cloud is calculated, and the point cloud is uniformly projected on the horizontal plane based on the resolution to generate a BEV image;

[0041] A pairing module is configured to extract key points and establish a correspondence between the key points in a source BEV image and a target BEV image generated by a source point cloud and a target point cloud respectively, and establish a correspondence between key points of the source point cloud and the target point cloud based on the correspondence between the key points of the source BEV image and the target BEV image.

[0042] An optimization module is configured to obtain an initial value of a transformation matrix of the source point cloud and the target point cloud based on the correspondence between the key points of the source point cloud and the target point cloud, and iteratively optimize the transformation matrix.

[0043] A third aspect of an embodiment of the present disclosure provides a point cloud registration device, comprising a memory and a processor,

[0044] The memory is configured to store a computer program;

[0045] The processor is configured to implement the method of the first aspect of the present disclosure when executing the computer program.

[0046] In summary, the point cloud registration method, device and equipment provided by the embodiments of the present disclosure convert the 3D point cloud into a 2D bird's eye view (BEV) image, and utilize key point matching to realize point cloud registration in a large-scale scene, which not only significantly improves the registration accuracy, but also greatly reduces the computational complexity. The traditional method has a large computational burden in large-scale point cloud registration, is strongly dependent on the design of descriptors, and has limited generalization ability. The method of the present disclosure overcomes these limitations through dimension reduction processing, so that efficient and robust registration results are still maintained in different scenes. In addition, the RANSAC algorithm is used to preprocess the point cloud to align the ground plane with the XOY plane, which simplifies the processing flow of the point cloud and provides a reliable benchmark for the unification of multi-view data. Overall, the point cloud registration method proposed in the present disclosure performs well in terms of accuracy, efficiency, generalization ability and processing simplification, and is particularly suitable for large-scale scene point cloud registration tasks with limited overlap. BRIEF DESCRIPTION OF DRAWINGS

[0047] The features and advantages of the present disclosure will be more clearly understood through reference to the following drawings, which are presented as illustrative and should not be construed as limiting the present disclosure, in which:

[0048] Figure 1 is a schematic diagram of a computer system to which the present disclosure is applicable;

[0049] Figure 2 is a flowchart of a point cloud registration method according to some embodiments of the present disclosure;

[0050] Figure 3 is laser point cloud data collected from the ground and aerial unmanned platforms respectively in an embodiment of the present disclosure;

[0051] Figure 4 is a BEV image generated from the point cloud data shown in Figure 3 ;

[0052] Figure 5 is a schematic diagram of key point extraction and matching in the BEV image shown in Figure 4 ;

[0053] Figure 6 is the corresponding relationship of the three-dimensional point cloud established based on the corresponding relationship of the key points of the BEV image shown in Figure 5 ;

[0054] Figure 7 is the result of point cloud registration shown in Figure 3 ;

[0055] Figure 8 is the point cloud registration result of an embodiment of the present disclosureFigure 2 Performance comparison results of the method described in S210-S230 and other point cloud registration algorithms on the GrAco dataset;

[0056] Figure 9 is a schematic diagram of a point cloud registration device according to some embodiments of the present disclosure;

[0057] Figure 10 is a schematic diagram of a point cloud registration device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0058] In the following detailed description, numerous specific details are set forth by way of examples in order to provide a thorough understanding of the relevant teachings. However, it should be apparent that the present disclosure can be practiced without such details. In other instances, well known methods, procedures, components, and network architectures have been described in detail and decided not to be figured in order to avoid obscuring aspects of the present disclosure.

[0059] It should be understood that when a device, unit, or module is referred to as being "on", "connected to", or "coupled to" another device, unit, or module, it can be directly on, connected, or coupled to the other device, unit, or module, or intervening devices, units, or modules can be present, unless the context clearly dictates otherwise. For example, as used in the present disclosure, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0060] The terminology used in the present disclosure is only for the purpose of describing particular embodiments and is not intended to be limiting of the present disclosure. As used in the description of the present disclosure and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0061] These and other features and characteristics of the present disclosure, as well as the methods of operation and functions of the related elements of the structure and the combination of parts and economies of manufacture, will become more apparent upon consideration of the following description and the appended claims with reference to the accompanying drawings, all of which form a part of this specification. It is to be expressly understood, however, that the drawings are for the purpose of illustration and description only and are not intended as a definition of the limits of the present disclosure. As such, it should be readily appreciated that the drawings are not necessarily drawn to scale of the embodiments thereof.

[0062] Various structural diagrams are used in the present disclosure to illustrate various variations according to embodiments of the present disclosure. It should be understood that the foregoing or the following structures are not intended to limit the present disclosure. The scope of protection of the present disclosure is subject to the claims.

[0063] Figure 1 is a schematic diagram of a computer system to which the present disclosure is applicable. Figure 1 The computer system shown includes a point cloud registration server connected with multiple laser radars, which acquire point cloud data of a target area from multiple angles, and the point cloud registration server performs point cloud registration on the point cloud data acquired by different laser radars.

[0064] In some embodiments of the present disclosure, the target area covers a very wide area, up to thousands of square meters. The laser radars can be deployed on different platforms to collect point clouds of the target area from different angles and paths, such as terrestrial laser scanning (TLS) point clouds formed by laser radars deployed on ground platforms and aerial laser scanning (ALS) point clouds formed by laser radars deployed on air platforms. The point cloud registration server can be any one of a single computer, a cluster, or a distributed server.

[0065] Due to the large data size of TLS and ALS point clouds, which are usually composed of millions of points and cover an area of thousands of square meters, and due to the different collection angles and paths, the overlap between the two is often small. Current point cloud registration methods either rely on handcrafted feature descriptors or are based on learning, where the method relying on handcrafted feature descriptors usually relies on specific structural features and is sensitive to scene complexity; and the learning-based method has low computational efficiency and high resource consumption in large-scale scenes. Therefore, neither of them can be applied to large-scale low-overlap scenes.

[0066] In view of this, the present disclosure proposes a method suitable for point cloud registration in large-scale scenes, particularly for the integration of TLS and ALS data. The method converts 3D point clouds into 2D bird's eye view (BEV) images through modal conversion, and promotes correspondence estimation through key point matching in 2D images. This BEV view can capture more overlapping information, thereby significantly improving the accuracy of registration in scenes with limited overlap. By using 2D BEV images for key point extraction and matching, the method greatly reduces the computational burden in large-scale scene tasks involving millions of points. In addition, the method breaks through the limitations of traditional handcrafted descriptor methods in descriptor design, while overcoming the challenges of learning-based methods in generalization ability, providing a more efficient and robust solution for point cloud registration in large-scale scenes.

[0067] Figure 2is a flowchart of a point cloud registration method according to some embodiments of the present disclosure. In some embodiments, the point cloud registration method is performed by a point cloud registration server as shown in Figure 1 includes the following steps:

[0068] S210, respectively collecting point cloud data of a target region from multiple perspectives, calculating the resolution of the point cloud in each of the point cloud data, uniformly projecting the point cloud on the ground plane based on the resolution, and generating a BEV image.

[0069] The present disclosure uniformly projects the point cloud on the ground plane based on the resolution to generate a BEV image. Resolution represents the number of pixels per inch in computer graphics, and in the present disclosure, resolution represents the number of laser points per inch. This significantly affects the quality of the BEV image. Too low a resolution can result in loss of point cloud details due to compression, while too high a resolution can produce holes that destroy the original point cloud structure. Given the huge difference in density between TLS and ALS point clouds, the present disclosure designs a resolution parameter RES that adapts to different point cloud densities.

[0070]

[0071] The resolution of the point cloud P is calculated by the above formula, where RES represents the resolution of the point cloud, N represents the number of points in the point cloud P, γ is a hyperparameter, x max , x min , y max , y min represent the maximum and minimum values of the point cloud P in the X and Y directions, and the denominator is the area occupied by P.

[0072] To facilitate the generation of the BEV image, the point cloud is first rotated so that the ground plane is parallel to the XOY plane. Specifically, the present disclosure uses the RANSAC algorithm to extract the ground plane in the point cloud, and calculates the rotation matrix according to the angle between the ground plane and the XOY plane. After this rotation, projecting the point cloud onto the XOY plane is equivalent to projecting it onto the ground plane. This operation not only simplifies the subsequent generation process of the BEV image, but also provides a unified reference framework for point clouds acquired at different heights, so that more accurate registration can be performed.

[0073] The present disclosure optimizes the continuity of detail expression and overall structure by adjusting the point-to-point distance of the point cloud. For high-density point clouds, the present disclosure increases the point-to-point distance to make the details more clear and distinguishable; while for low-density point clouds, the point-to-point distance is reduced to maintain the coherence of the outline, thereby effectively reducing the generation of holes and the identification of false noise. This strategy ensures accurate point cloud expression and registration results in different density areas.

[0074] To convert a 3D point cloud to a BEV image, the present disclosure first establishes a mapping from the point cloud coordinate system to the pixel coordinate system. For a point cloud P with N points, define

[0075]

[0076] where D ij represents the set of points projected to the pixel at pixel coordinate system coordinate (i,j), P is a point cloud with N points, and the X and Y coordinates of the kth point p k are x k and y k , respectively, and x min and y min are the minimum values in the X and Y directions of the point cloud coordinate system, respectively.

[0077] Next, define H ij as the maximum height value of the points in D ij , which can be expressed as:

[0078]

[0079] A BEV image matrix G is generated using H, where each element G ij corresponds to the grayscale value of the pixel at coordinate (i,j). This can be expressed as:

[0080]

[0081] where H ij is the maximum height value of the points in D ij , G ij is the grayscale value of the pixel at coordinate (i,j) in the BEV image matrix G, D ij represents the set of points projected to the pixel at pixel coordinate system coordinate (i,j), z max and z min are the maximum and minimum values in the Z direction of the point cloud coordinate system, respectively, and the height value of the kth point p k is z k .

[0082] The present invention uses image enhancement techniques to improve the accuracy of key point extraction from images. Two high-pass filters of different intensities are used to enhance the high-frequency content in the image, thereby improving the distinguishability of structural elements such as buildings, roads, and vegetation. The process can be summarized as follows:

[0083]

[0084] where, is the enhanced image matrix, w1 represents the convolution kernel for edge extraction, w2 represents the convolution kernel for sharpening details, and ★ represents the convolution operation.

[0085] In one embodiment of the present disclosure, the laser point cloud data collected from the ground and the aerial unmanned platform respectively to be registered is as shown in Figure 3 Figure 4 is a BEV image generated from the point cloud data as shown in Figure 1

[0086] S220, in the source BEV image and the target BEV image generated respectively from the source point cloud and the target point cloud, key points are extracted and a corresponding relationship between the key points is established, and a corresponding relationship between the key points of the source point cloud and the target point cloud is established based on the corresponding relationship between the key points of the source BEV image and the target BEV image.

[0087] Some embodiments of the present disclosure extract key points of the BEV image by using SuperPoint, and then obtain 2D point correspondence through the image key point matching framework LightGlue. For the case that the overlap between two point clouds is too low, the FOCUS module is designed to improve the registration accuracy. If the proportion of the overlapping area in the point cloud is lower than a predefined threshold θ, key point extraction and matching will be performed again on the neighborhood of the matching points obtained from the first matching.

[0088] Key point extraction and matching in the BEV image as shown in Figure 2 Figure 5

[0089] According to the 2D key point matching result, the present disclosure remaps the key points and their corresponding relationship to the 3D point cloud, and the mapping relationship is as shown below:

[0090]

[0091] wherein RES represents the resolution of the point cloud, X, Y and Z are the coordinates of the point cloud points in the three-dimensional coordinate system, G ij corresponds to the pixel gray value of the BEV image at coordinates (i, j), x min and y min are the minimum values of the point cloud coordinate system X and Y directions respectively, z max and z min are the maximum and minimum values of the point cloud coordinate system Z direction respectively.

[0092] The corresponding relationship of the three-dimensional point cloud established based on the corresponding relationship of the key points of the BEV image as shown in Figure 5 Figure 6

[0093] S230, based on the corresponding relationship of the key points of the source point cloud and the target point cloud, an initial value of a transformation matrix of the source point cloud and the target point cloud is obtained, and the transformation matrix is iteratively optimized. ​​​​​​

[0094] This disclosure uses the correspondence between key points of the source point cloud and the target point cloud as the correspondence between the source point cloud and the target point cloud, and uses SVD to estimate the transformation matrix between the point clouds. To improve the estimation accuracy, an iterative optimization strategy is adopted. Because ICP is sensitive to initial values, SVD is used to estimate the initial values ​​first. Specifically, by iteratively executing the SVD process, outlier correspondences can be effectively filtered out. After the initial estimation of the transformation matrix, it is used as the initial value input into the ICP algorithm to achieve more refined point cloud registration.

[0095] based on Figure 2 The method described in S210-S230 (hereinafter referred to as "this method") is for Figure 3 The point cloud registration results shown are as follows: Figure 7 As shown.

[0096] An embodiment of this disclosure validates the effectiveness of the proposed method on the GrAco dataset. The experiment compares this method with four high-performing point cloud registration algorithms: FPFH+TEASER, PLADE, Predator, and GeoTransformer. Using "G" to represent the ground sequence and "A" to represent the aerial sequence, our experiment focuses on combinations of eight AGs, three GGs, and three AAs from the source and target point clouds. Rotation error is used... Translation error The root mean square distance (RMSD) is used to measure the accuracy of the registration results. Experimental results are as follows: Figure 8 As shown:

[0097] Figure 8 The results show that our proposed method performs well across all types of point cloud combinations. It maintains low errors in both complex AG combinations and relatively simple AA and GG combinations. This indicates that our method exhibits high robustness and versatility in handling large-scale point cloud registration with limited overlap.

[0098] Figure 9 This is a schematic diagram of a point cloud registration apparatus according to some embodiments of the present disclosure. Figure 9 As shown, the point cloud registration device 900 includes a projection module 910, a pairing module 920, and an optimization module 930. In some embodiments of this disclosure, the point cloud registration function is performed by... Figure 1 The point cloud registration server shown is executing the command. Among other things:

[0099] The projection module 910 is used to collect point cloud data of the target area from multiple perspectives, calculate the resolution of the point cloud in each point cloud data, and project the point cloud uniformly onto the ground plane based on the resolution to generate a BEV image.

[0100] The pairing module 920 is used to extract key points and establish a correspondence between the key points in the source BEV image and the target BEV image generated from the source point cloud and the target point cloud respectively, and to establish a correspondence between the key points of the source point cloud and the target point cloud based on the correspondence between the key points of the source BEV image and the target BEV image.

[0101] The optimization module 930 is used to obtain the initial value of the transformation matrix of the source point cloud and the target point cloud based on the correspondence between the key points of the source point cloud and the target point cloud, and to iteratively optimize the transformation matrix.

[0102] One embodiment of this disclosure provides a point cloud registration device. For example... Figure 10 As shown, the point cloud registration device 1000 includes a memory 1020 and a processor 1010. The memory 1020 is used to store computer programs; the processor 1010 is used to implement... when the computer program is executed. Figure 2 The method described in S210-S230.

[0103] In summary, the point cloud registration method, apparatus, and device provided in the embodiments of this disclosure, by converting 3D point clouds into 2D bird's-eye view (BEV) images and utilizing key point matching to achieve point cloud registration in large-scale scenes, not only significantly improves registration accuracy but also greatly reduces computational complexity. Traditional methods in large-scale point cloud registration have high computational burden, strong dependence on descriptor design, and limited generalization ability. The method of this disclosure overcomes these limitations through dimensionality reduction, maintaining efficient and robust registration results in different scenarios. Furthermore, by preprocessing the point cloud using the RANSAC algorithm to align the ground plane with the XOY plane, the point cloud processing flow is simplified, providing a reliable benchmark for the unification of multi-view data. Overall, the point cloud registration method proposed in this disclosure performs superiorly in terms of accuracy, efficiency, generalization ability, and processing simplification, and is particularly suitable for point cloud registration tasks in large-scale scenes with limited overlap. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and modules described above can be referred to the corresponding descriptions in the foregoing apparatus embodiments, and will not be repeated here.

[0104] Although the subject matter described herein is provided in the general context of computer-executable instructions of a program module being executed by a computer system on a computing device, those skilled in the art will recognize that programs modules can be implemented in cooperation with other programs modules or not, and computer system or systems can be implemented as a combination of computing devices executing programs modules, as at least partially embodied in a microprocessor-based or in programmable consumer electronics, as at least partially embodied in a minicomputer, mainframe computer, and the like. Still further, those skilled in the art will appreciate that one or more computer systems with associated computer program modules / members can be deployed, while software modules implemented as computer program modules can be stored on a computer usable medium almost anywhere on a given computer system — from a hard disk to a computer data storage media. Also, those skilled in the art will appreciate that the subject matter described herein can be implemented as a computer program product, i.e., a computer program tangibly embodied in a computer readable storage medium.

[0105] Those skilled in the art can realize also that the exemplary units and method steps described in connection with the embodiments disclosed herein can be carried out by electronic hardware, computer software, or combinations of both. Whether such functionality is implemented in hardware or software depends on the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0106] It is to be understood that the above specific embodiments of the present disclosure are merely illustrative of the principles of the present disclosure and are not intended to limit the scope of the present disclosure. Any modification, equivalent replacement, improvement, and the like of any part of the present disclosure made without departing from the spirit and scope of the present disclosure should be included in the scope of the present disclosure. In addition, the appended claims of the present disclosure are intended to cover all variations and modifications falling within the scope and boundary of the appended claims or the equivalent form of such scope and boundary.

Claims

1. A point cloud registration method, characterized in that, include: Point cloud data of the target area are collected from multiple perspectives. In each point cloud data, the resolution of the point cloud is calculated. Based on the resolution, the point cloud is uniformly projected onto the ground plane to generate a BEV image. In the source BEV image and target BEV image generated from the source point cloud and target point cloud respectively, key points are extracted and the correspondence between the key points is established. Based on the correspondence between the key points of the source BEV image and the target BEV image, the correspondence between the key points of the source point cloud and the target point cloud is established. Based on the correspondence between the key points of the source point cloud and the target point cloud, the initial values ​​of the transformation matrix of the source point cloud and the target point cloud are obtained, and the transformation matrix is ​​iteratively optimized. The resolution of the computed point cloud includes: Where RES represents the resolution of the point cloud, N represents the number of points in the point cloud, γ is a hyperparameter, and x max ,x min ,y max ,y min These represent the maximum and minimum values ​​of the point cloud in the X and Y directions, respectively; The step of uniformly projecting the point cloud onto the ground plane based on the resolution includes: Extract the ground plane from the point cloud and obtain the angle between the ground plane and the XOY plane; Rotate the point cloud based on the included angle so that the ground plane is parallel to the XOY plane; Adjust the point-to-point distance in the point cloud so that the point cloud is uniformly projected onto the XOY plane based on the resolution; The generated BEV image includes: Establish a mapping from the point cloud coordinate system to the pixel coordinate system: Among them, D ij Let P be a point cloud with N points, where the point at coordinate (i,j) is the point projected onto the pixel in the pixel coordinate system. k The X and Y coordinates are x k and y k x min and y min These are the minimum values ​​in the X and Y directions, respectively, in the point cloud coordinate system; Based on D ij Maximum height H at the midpoint ij Generate a BEV image matrix G, where each element G ij The gray value corresponding to the pixel at coordinates (i,j) Among them, H ij D ij The maximum height value of the midpoint, G ij Let D be the gray value of the pixel at coordinates (i,j) in the BEV image matrix G. ij This represents the set of points projected onto the pixel at coordinates (i,j) in the pixel coordinate system. max and z min These are the maximum and minimum values ​​in the Z direction of the point cloud coordinate system, respectively, for the k-th point p. k The height value is z k .

2. The method according to claim 1, characterized in that, After generating the BEV image, the process further includes: Two high-pass filters of different strengths are used to enhance the high-frequency content in an image. The filtering operation is represented as follows: in, It is an image enhancement matrix, where w1 represents the convolution kernel used to extract edges, w2 represents the convolution kernel used to sharpen details, and ★ represents the convolution operation.

3. The method according to claim 2, characterized in that, The process of extracting key points and establishing the correspondence between the key points includes: Extracting key points from BEV images using SuperPoint; The key points of different BEV images generated from different point clouds of the target region are paired using the image key point matching framework LightGlue. If the number of matching key points with a pairing relationship is less than a preset threshold, then key point extraction and matching are performed again on the neighborhood of the matching key points.

4. The method according to claim 3, characterized in that, The step of establishing the correspondence between the key points of the source point cloud and the target point cloud based on the correspondence between the key points of the source BEV image and the target BEV image includes: The matching key points of the source BEV image and the target BEV image are mapped to the source point cloud and the target point cloud, respectively, to obtain point cloud points with the corresponding relationship, wherein the mapping relationship is: RES represents the resolution of the point cloud, X, Y, and Z are the coordinates of the point cloud points in the three-dimensional coordinate system, and G... ij The pixel grayscale value corresponding to the BEV image at coordinates (i,j), x min and y min These are the minimum values ​​in the X and Y directions of the point cloud coordinate system, respectively, and z. max and z min These are the maximum and minimum values ​​in the Z direction of the point cloud coordinate system, respectively.

5. The method according to claim 4, characterized in that, The iterative optimization of the transformation matrix includes: The correspondence of the point cloud represented by the transformation matrix is ​​input into the SVD process and iteratively executed to update the transformation matrix until the registration of the source point cloud and the target point cloud based on the transformation matrix reaches a preset standard. The transformation matrix is ​​then input into the ICP algorithm for further registration.

6. A point cloud registration device, characterized in that, include: The projection module is used to collect point cloud data of the target area from multiple perspectives, calculate the resolution of the point cloud in each point cloud data, and project the point cloud uniformly onto the ground plane based on the resolution to generate a BEV image. The pairing module is used to extract key points from the source BEV image and target BEV image generated from the source point cloud and target point cloud respectively, and establish the correspondence between the key points. Based on the correspondence between the key points of the source BEV image and the target BEV image, the correspondence between the key points of the source point cloud and the target point cloud is established. An optimization module is used to obtain the initial value of the transformation matrix of the source point cloud and the target point cloud based on the correspondence between the key points of the source point cloud and the target point cloud, and to iteratively optimize the transformation matrix; The resolution of the computed point cloud includes: Where RES represents the resolution of the point cloud, N represents the number of points in the point cloud, γ is a hyperparameter, and x max ,x min ,y max ,y min These represent the maximum and minimum values ​​of the point cloud in the X and Y directions, respectively; The step of uniformly projecting the point cloud onto the ground plane based on the resolution includes: Extract the ground plane from the point cloud and obtain the angle between the ground plane and the XOY plane; Rotate the point cloud based on the included angle so that the ground plane is parallel to the XOY plane; Adjust the point-to-point distance in the point cloud so that the point cloud is uniformly projected onto the XOY plane based on the resolution; The generated BEV image includes: Establish a mapping from the point cloud coordinate system to the pixel coordinate system: Among them, D ij Let P be a point cloud with N points, where the point at coordinate (i,j) is the point projected onto the pixel in the pixel coordinate system. k The X and Y coordinates are x k and y k x min and y min These are the minimum values ​​in the X and Y directions, respectively, in the point cloud coordinate system; Based on D ij Maximum height H at the midpoint ij Generate a BEV image matrix G, where each element G ij The gray value corresponding to the pixel at coordinates (i,j) Among them, H ij D ij The maximum height value of the midpoint, G ij Let D be the gray value of the pixel at coordinates (i,j) in the BEV image matrix G. ij This represents the set of points projected onto the pixel at coordinates (i,j) in the pixel coordinate system. max and z min These are the maximum and minimum values ​​in the Z direction of the point cloud coordinate system, respectively, for the k-th point p. k The height value is z k .

7. A point cloud registration device, characterized in that, Including memory and processor, The memory is used to store computer programs; The processor is configured to implement the method of any one of claims 1-5 when executing the computer program.

Images