Heterogeneous point cloud global positioning method based on multi-sector loss strategy

Abstract

The invention discloses a heterogeneous point cloud global positioning method based on a multi-sector loss strategy, and the method comprises the following steps: firstly, dividing a laser radar point cloud from any model according to sectors, and projecting the point cloud to a radial-elevation plane of a cylindrical coordinate system, thereby obtaining an occupation feature matrix; secondly, sector similarity between the query frame and the map frame is calculated through a sliding window and a superposition loss strategy, and candidate queues are generated according to ranking of the maximum similarity; and finally, taking a sector similar position result as initial rotation, and screening out potential error retrieval through a voxel overlapping similarity scoring mechanism and a GICP technology, thereby outputting an accurate six-degree-of-freedom pose estimation result. According to the method, challenges brought by different field angles, different resolutions and different scanning methods in current heterogeneous laser radar global positioning are effectively overcome, scene details are reserved by using contour features in sectors through a feature extraction strategy based on sector division, and the precision and robustness of global positioning are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser radar global positioning, and in particular relates to a heterogeneous point cloud global positioning method based on a multi-sector loss strategy. Background Art

[0002] One-shot global localization plays a key role in most robotics applications, offering significant advantages during initialization and relocalization. However, current heterogeneous LiDAR-based one-shot global localization approaches struggle to achieve robust and fast global localization without prior poses, facing challenges such as varying fields of view (FoV), resolutions, and scanning methods.

[0003] Existing one-shot global localization methods have limitations when dealing with complex environments. They are all based on high-resolution, mechanically rotating lidars, making them difficult to scale effectively to devices like solid-state lidars with smaller fields of view, and thus unable to achieve stable and robust global localization. Therefore, a global localization method for heterogeneous point clouds based on a multi-sector loss strategy is proposed to address these issues. Summary of the Invention

[0004] To solve the above problems, the present invention discloses a global positioning method for heterogeneous point clouds based on a multi-sector loss strategy. By analyzing the characteristics of lidar sensor data, the contour shape of the object scanned in each sector is feature encoded, and adaptive sliding window technology is used to achieve unified encoding from narrow field of view to wide field of view, thereby improving the accuracy and robustness of global positioning.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] The global positioning method of heterogeneous point cloud based on multi-sector loss strategy includes the following steps:

[0007] S1. Divide the point cloud from any model of LiDAR into sectors and project it onto the radial-elevation plane of a cylindrical coordinate system to obtain an occupancy feature matrix;

[0008] S2. Input the flattened and reduced-dimensional feature matrix into the system, calculate the sector similarity between the query frame and the map frame through the sliding window and overlay loss strategy, and generate a candidate queue based on the maximum similarity sorting.

[0009] S3. The sector similarity position results are used as the initial rotation estimation, and potential erroneous retrievals are screened out through the voxel overlap similarity scoring mechanism and GICP technology, thereby outputting an accurate 6-DOF pose estimation result.

[0010] The specific steps are:

[0011] Step S1. Divide the point cloud from any model of LiDAR into sectors and project it onto the radial-elevation plane of the cylindrical coordinate system to obtain the occupancy feature matrix

[0012] When processing the input point cloud Q, the present invention eliminates the inherent variation in horizontal and vertical FoV between different LiDAR models through cylindrical coordinate conversion. Unlike the spherical coordinate system adopted in existing methods, the cylindrical method of the present invention provides excellent adaptability by decomposing the LiDAR beam profile in the axial direction while maintaining the radial resolution. For each point p in the point cloud Q, initially defined as p = {x, y, z} in Cartesian coordinates, the present invention performs a conversion to cylindrical coordinates p = {r, θ, z} according to formula (1):

[0013]

[0014] Our method employs a divide-and-conquer strategy to process point clouds through sectorization. First, we discretize the point cloud within each sector using voxelization. We then project these 3D voxels onto the radial-elevation plane of the corresponding sector, generating an N-layer BEV feature map through hierarchical projection. This representation, inherently rotationally invariant through polar coordinate decomposition, effectively preserves the scene contours within each sector.

[0015] Step S2. To further enhance feature compatibility across heterogeneous point clouds, the present invention reconstructs the N × R × H three-dimensional feature matrix into a two-dimensional N × (R·H) feature descriptor through tensor expansion. This transformation significantly reduces computational complexity while maintaining feature distinguishability, ultimately achieving robust cross-sensor localization. Furthermore, compared to other descriptor construction methods, the proposed scheme ensures the independence of sector features, providing a new perspective on solving the problem of feature unification for heterogeneous point clouds with different fields of view.

[0016] In order to solve the dimensionality mismatch problem of heterogeneous lidar feature descriptors, this paper proposes an overlapping loss matching mechanism based on sliding windows. Consider a map frame feature matrix F constructed by an Ouster-128 lidar. O (The dimension is N O ×(R·H)), as a reference sequence, and a feature matrix F constructed by the Livox-Mid70 lidar L (The dimension is N L ×(R·H)), as the query sequence (where N O >N L The present invention realizes heterogeneous point cloud matching by constructing an extended matrix with the dimension of (N O +N L -1)×(R·H), define a size of NL ×(R·H) sliding window At position k. Initially, the basic similarity can be formulated as:

[0017]

[0018] Each sliding window position k (1≤k≤N O -N L +1) similarity score S m (k) quantifies the degree of alignment of the scene contours between the map point cloud and the query point cloud in different sectors. In order to further improve the discrimination ability, the present invention uses an overlap loss function to improve this similarity metric:

[0019]

[0020] This overlapping loss strategy implements a weighting mechanism: as shown in Equation (10), when the feature descriptor does not exist at the same location (all zero state), no score is assigned; for the overlapping activation region (both features exist), positive feedback (s a >0); for conflicting activated regions (only one feature active), negative feedback is applied (s b <0). This method realizes the dynamic weighted feature activation mode through formula (9), which effectively enhances the matching robustness of scene structure features.

[0021]

[0022] The optimal alignment position is determined by equation (11), which determines the sliding window position that produces the maximum similarity score. From the obtained complete window similarity distribution, the present invention extracts the peak position k * As the optimal matching result of the present invention, this directly corresponds to the initial rotation angle estimation between heterogeneous point clouds. This method effectively solves the dimensionality mismatch problem and significantly improves the reliability of cross-sensor matching by finely evaluating feature activation patterns. After completing the traversal of all map frames, the present invention further selects the top Num with the highest similarity score. K maps to form candidate frame sequences as input to subsequent processing modules.

[0023] Step S3. The preliminary retrieval module only provides 2D position identification and initial rotation estimation. Projecting the point cloud onto the radial-elevation plane inevitably results in a loss of 3D structural information, which can lead to mismatches. To overcome these limitations, the present invention designs a verification module based on voxel spatial alignment. This module effectively compensates for the limitations of the feature matching stage through precise 3D geometric consistency checks, ensuring comprehensive and accurate pose estimation for reliable localization in complex environments.

[0024] First, the present invention performs a multi-dimensional multi-dimensional multi-map transformation on the candidate map point cloud P in voxel space. O and query point cloud P L Discretization is performed to standardize the spatial resolution, effectively addressing the challenges posed by heterogeneous point clouds with different resolutions and scanning modes. Subsequently, the present invention performs GICP based on the initial rotation pose provided by the preliminary retrieval stage. Finally, the present invention evaluates the matching quality by calculating the overlap metric between the registered point clouds:

[0025]

[0026] Where P and Q represent the voxelized candidate points and query point cloud respectively, N is the number of points in the point set P, is the indicator function, when point p i There is at least one point q within distance ∈ j In the point set Q, the function is equal to 1. When the overlap measure exceeds the threshold Thre O When , the match is considered valid and the 6-DOF pose optimized by GICP is output; otherwise, the re-retrieval mechanism is triggered to check the next candidate frame in the queue.

[0027] The beneficial effects of the present invention are:

[0028] The present invention demonstrates superior performance in the three heterogeneous lidar scenarios of the eight sequences tested: the median of its translation error is generally maintained below 0.5, and the span of the box and whisker is significantly smaller than that of the comparison methods (such as BTC and BEVPlace). This not only reflects the high positioning accuracy of the algorithm, but also shows that its results are stable and consistent, fully confirming the robustness and reliability of the present invention in translation error control. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flowchart of the implementation of the heterogeneous point cloud global positioning method based on the multi-sector loss strategy of the present invention;

[0030] Figure 2 It is a radar chart showing the performance of the method of the present invention compared with other open source methods on various data sets;

[0031] Figure 3 It is a time-consuming analysis diagram of each module in the method described in the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0033] As shown in the figure, the heterogeneous point cloud global positioning method based on the multi-sector loss strategy of the present invention has the following specific steps:

[0034] Step 1: The method for global positioning of heterogeneous point clouds based on a multi-sector loss strategy according to claim 1, wherein the point cloud from any model of LiDAR is divided into sectors and projected onto a radial-elevation plane of a cylindrical coordinate system to obtain an occupancy feature matrix; specifically, the following steps are performed:

[0035] When processing the input point cloud Q, the present invention eliminates the inherent variation in horizontal and vertical FoV between different LiDAR models through cylindrical coordinate conversion. Unlike the spherical coordinate system adopted in existing methods, the cylindrical method of the present invention provides excellent adaptability by decomposing the LiDAR beam profile through axial decomposition while maintaining the radial resolution. For each point p in the point cloud Q, initially defined as p = {x, y, z} in Cartesian coordinates, the present invention performs a conversion to cylindrical coordinates p = {r, θ, z} according to formula (13):

[0036]

[0037] Our method employs a divide-and-conquer strategy to process point clouds through sectorization. First, we discretize the point cloud within each sector using voxelization. We then project these 3D voxels onto the radial-elevation plane of the corresponding sector, generating an N-layer BEV feature map through hierarchical projection. This representation, inherently rotationally invariant through polar coordinate decomposition, effectively preserves the scene contours within each sector.

[0038] Step 2: To further enhance feature compatibility across heterogeneous point clouds, the present invention reconstructs the N × R × H three-dimensional feature matrix into a two-dimensional N × (R·H) feature descriptor through tensor expansion. This transformation significantly reduces computational complexity while maintaining feature distinguishability, ultimately achieving robust cross-sensor localization. Furthermore, compared to other descriptor construction methods, the proposed scheme ensures the independence of sector features, providing a new perspective on solving the problem of feature unification for heterogeneous point clouds with different fields of view.

[0039] In addition, in order to solve the dimensionality mismatch problem of heterogeneous lidar feature descriptors, this paper proposes an overlapping loss matching mechanism based on sliding windows. Consider a map frame feature matrix F constructed by an Ouster-128 lidar. O (The dimension is N O ×(R·H)), as a reference sequence, and a feature matrix F constructed by the Livox-Mid70 lidar L (The dimension is N L ×(R·H)), as the query sequence (where NO >N L The present invention realizes heterogeneous point cloud matching by constructing an extended matrix with the dimension of (N O +N L -1)×(R·H), define a size of N L ×(R·H) sliding window At position k. Initially, the basic similarity can be formulated as:

[0040]

[0041] Each sliding window position k (1≤k≤N O -N L +1) similarity score S m (k) quantifies the degree of alignment of the scene contours between the map point cloud and the query point cloud in different sectors. In order to further improve the discrimination ability, the present invention uses an overlap loss function to improve this similarity metric:

[0042]

[0043] This overlapping loss strategy implements a weighting mechanism: as shown in Equation (16), when the feature descriptor does not exist at the same location (all zero state), no score is assigned; for the overlapping activation region (both features exist), positive feedback (s a >0); for conflicting activated regions (only one feature active), negative feedback is applied (s b <0). This method realizes the dynamic weighted feature activation mode through formula (15), which effectively enhances the matching robustness of scene structure features.

[0044]

[0045] The optimal alignment position is determined by equation (17), which determines the sliding window position that produces the maximum similarity score. From the obtained complete window similarity distribution, the present invention extracts the peak position k * As the optimal matching result of the present invention, this directly corresponds to the initial rotation angle estimation between heterogeneous point clouds. This method effectively solves the dimensionality mismatch problem and significantly improves the reliability of cross-sensor matching by finely evaluating feature activation patterns. After completing the traversal of all map frames, the present invention further selects the top Num with the highest similarity score. K maps to form candidate frame sequences as input to subsequent processing modules.

[0046] Step 3: The preliminary retrieval module only provides 2D position identification and initial rotation estimation. Projecting the point cloud onto the radial-elevation plane inevitably results in a loss of 3D structural information—potentially leading to mismatches. To overcome these limitations, the present invention meticulously designs a verification module based on voxel spatial alignment. This module effectively compensates for the limitations of the feature matching stage through precise 3D geometric consistency checks, ensuring comprehensive and accurate pose estimation for reliable localization in complex environments.

[0047] First, the present invention performs a multi-dimensional multi-dimensional multi-map transformation on the candidate map point cloud P in voxel space. O and query point cloud P L Discretization is performed to standardize the spatial resolution, effectively addressing the challenges posed by heterogeneous point clouds with different resolutions and scanning modes. Subsequently, the present invention performs GICP based on the initial rotation pose provided by the preliminary retrieval stage. Finally, the present invention evaluates the matching quality by calculating the overlap metric between the registered point clouds:

[0048]

[0049] Where P and Q represent the voxelized candidate points and query point cloud respectively, N is the number of points in the point set P, is the indicator function, when point p i There is at least one point q within distance ∈ j In the point set Q, the function is equal to 1. When the overlap measure exceeds the threshold Thre O When , the match is considered valid and the 6-DOF pose optimized by GICP is output; otherwise, the re-retrieval mechanism is triggered to check the next candidate frame in the queue.

[0050] It should be noted that the above content merely illustrates the technical idea of the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A global positioning method for heterogeneous point clouds based on a multi-sector loss strategy, characterized by: include: Step 1: Divide the point cloud from any model of LiDAR into sectors and project it onto the radial-elevation plane of the cylindrical coordinate system to obtain the occupancy feature matrix; Step 2: Input the flattened and reduced feature matrix into the system, calculate the sector similarity between the query frame and the map frame through the sliding window and overlay loss strategy, and generate a candidate queue based on the maximum similarity ranking; Step 3: Use the sector similarity position result as the initial rotation estimate, and screen out potential false retrievals through the voxel overlap similarity scoring mechanism and GICP technology, thereby outputting an accurate 6-DOF pose estimation result.

2. The heterogeneous point cloud global positioning method based on multi-sector loss strategy according to claim 1, characterized in that: As described in step 1, the point cloud from any model of LiDAR is divided into sectors and projected onto the radial-elevation plane of the cylindrical coordinate system to obtain the occupancy feature matrix; the details are as follows: For each point p in the input point cloud Q, initially defined as p = {x, y, z} in Cartesian coordinates, a transformation to cylindrical coordinates p = {r, θ, z} is performed according to formula (1): A divide-and-conquer strategy is adopted to process point clouds by sectorization. First, the point cloud in each sector is discretized using voxelization. Then, these 3D voxels are projected onto the radial-elevation plane of the corresponding sector, and N-layer BEV feature maps are generated through hierarchical projection.

3. The heterogeneous point cloud global positioning method based on multi-sector loss strategy according to claim 2, characterized in that: Step 2 is as follows: In order to further enhance the feature compatibility between heterogeneous point clouds, the three-dimensional feature matrix of N×R×H is reconstructed into a two-dimensional feature descriptor of N×(R·H) through tensor expansion operation. In addition, in order to solve the dimensionality mismatch problem of heterogeneous lidar feature descriptors, an overlapping loss matching mechanism based on sliding window is proposed. Consider a map frame feature matrix F constructed by Ouster-128 lidar. O , the dimension is N O ×(R·H), as a reference sequence, and a feature matrix F constructed by the Livox-Mid70 lidar L , the dimension is N L ×(R·H), as the query sequence, where N O >N L ; Heterogeneous point cloud matching is achieved by constructing an extended matrix with the dimension (N O +N L -1)×(R·H), define a size of N L ×(R·H) sliding window At position k; initially, the basic similarity is formulated as: Each sliding window position k (1≤k≤N O -N L +1) similarity score S m (k) quantifies the degree of alignment of the scene outline between the map point cloud and the query point cloud in different sectors; to further improve the discrimination ability, an overlap loss function is used to improve this similarity measure: This overlapping loss strategy implements a weighting mechanism: as shown in Equation (4), when the feature descriptor does not exist at the same location, no score is assigned; for the overlapping activation areas, positive feedback (s a >0); for conflict-activated regions, negative feedback is applied (s b <0); This method realizes the dynamic weighted feature activation mode by calculating formula (3), which effectively enhances the matching robustness of scene structure features; The optimal alignment position is determined by equation (5), which determines the sliding window position that produces the maximum similarity score; from the obtained full window similarity distribution, the peak position k is extracted * As the optimal matching result, this directly corresponds to the initial rotation angle estimation between heterogeneous point clouds; after completing the traversal of all map frames, the top Num with the highest similarity score is further selected. K maps to form candidate frame sequences as input to subsequent processing modules.

4. The heterogeneous point cloud global positioning method based on multi-sector loss strategy according to claim 3, characterized in that: Step 3 is as follows: The preliminary retrieval module only provides 2D position identification and initial rotation estimation, and projecting the point cloud onto the radial-elevation plane inevitably leads to the loss of 3D structural information. To overcome these limitations, a verification module based on voxel spatial alignment is designed. This module compensates for the limitations of the feature matching stage through 3D geometric consistency checks, ensuring comprehensive and accurate pose estimation for reliable positioning in complex environments. First, the candidate map point cloud P is mapped in voxel space O and query point cloud P L Discretization is performed to standardize spatial resolution and effectively address the challenges posed by heterogeneous point clouds with different resolutions and scanning modes; Subsequently, GICP is performed based on the initial rotation pose provided by the preliminary retrieval stage; finally, the matching quality is evaluated by computing the overlap metric between the registered point clouds: Where P and Q represent the voxelized candidate points and query point cloud respectively, N is the number of points in the point set P, is the indicator function, when point p i There is at least one point q within distance ∈ j When the point set Q is in the set, the function is equal to 1; when the overlap measure exceeds the threshold Thre O When , the match is considered valid and the 6-DOF pose optimized by GICP is output; otherwise, the re-retrieval mechanism is triggered to check the next candidate frame in the queue.

Images