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

Abstract

The application discloses a kind of heterogeneous point cloud global positioning methods based on multi-sector loss strategy, comprising the following steps: first, from arbitrary type laser radar point cloud is divided according to sector, and is projected to radial-elevation plane of cylindrical coordinate system, to obtain the matrix of occupation characteristics;Second, the sector similarity between query frame and map frame is calculated by sliding window and superposition loss strategy, and candidate queue is generated according to maximum similarity sorting;Finally, sector similar position result is used as initial rotation, and potential false search is screened out by voxel overlap similarity scoring mechanism and GICP technology, to output accurate 6 degree of freedom pose estimation result;The method effectively overcomes the challenge brought by different field angles, different resolutions and different scanning methods in current heterogeneous laser radar global positioning, uses the profile feature in sector to retain scene details by the feature extraction strategy based on sector division, improves the accuracy and robustness of global positioning.

Description

Technical Field

[0001] This invention belongs to the field of lidar global positioning technology, specifically relating to a heterogeneous point cloud global positioning method based on a multi-sector loss strategy. Background Technology

[0002] One-time global localization plays a crucial role in most robotic applications, offering significant advantages during initialization and relocalization. However, current one-time global localization based on heterogeneous LiDAR still struggles to achieve robust and fast global localization capabilities when faced with challenges such as different field of view (FoV), different resolutions, and different scanning methods, especially without prior pose.

[0003] Existing one-time global localization methods have certain limitations when handling complex environments. They are all developed based on high-resolution mechanically rotating lidar, making it difficult to effectively extend to devices with smaller fields of view, such as solid-state lidar, and thus failing to achieve stable and robust global localization results. Therefore, this paper proposes a heterogeneous point cloud global localization method based on a multi-sector loss strategy to address these issues. Summary of the Invention

[0004] To address the aforementioned issues, this invention discloses a heterogeneous point cloud global localization method based on a multi-sector loss strategy. By analyzing the characteristics of lidar sensor data, the method encodes the feature shape of the object contour scanned in each sector and employs an adaptive sliding window technique to achieve unified encoding from narrow field of view to wide field of view, thereby improving the accuracy and robustness of global localization.

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

[0006] A heterogeneous point cloud global localization method based on a multi-sector loss strategy includes the following steps:

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

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

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

[0010] The specific steps are as follows:

[0011] Step S1. Divide the point cloud from any type 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, this invention eliminates the inherent variations in horizontal and vertical FoV between different LiDAR models through cylindrical coordinate transformation. Unlike the spherical coordinate system used in existing methods, the cylindrical method of this invention provides superior adaptability by maintaining the laser radar beam distribution characteristics with radial resolution through axial decomposition. For each point p in the point cloud Q, initially defined as p = {x, y, z} in Cartesian coordinates, this invention performs a transformation to cylindrical coordinates p = {r, θ, z} according to formula (1):

[0013]

[0014] The method of this invention employs a divide-and-conquer strategy, processing point clouds through sectorization. First, this invention discretizes the point cloud within each sector using voxelization. Then, this invention projects these 3D voxels onto the radial-elevation plane of the corresponding sector, generating an N-layer BEV feature map through hierarchical projection. This representation possesses inherent rotation invariance through polar coordinate decomposition, effectively preserving the scene contour features within each sector.

[0015] Step S2. To further enhance feature compatibility among heterogeneous point clouds, this 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 for solving the feature unification problem of heterogeneous point clouds with different field of view angles.

[0016] To address the dimensionality mismatch issue of feature descriptors from heterogeneous LiDAR systems, this invention proposes an overlap loss matching mechanism based on a sliding window. Consider a map frame feature matrix F constructed from an Ouster-128 LiDAR system. O (Dimension is N) O ×(R·H)), as a reference sequence, and a feature matrix F constructed from a Livox-Mid70 lidar. L (Dimension is N) L ×(R·H)), as the query sequence (where N is N) O >N L This invention achieves heterogeneous point cloud matching by constructing an extended matrix, the matrix of which has a dimension of (N). O +N L -1)×(R·H), define a size of NL Sliding window of ×(R·H) At position k, the basic similarity can initially be formulated as:

[0017]

[0018] Each sliding window position k (1≤k≤N) O -N L The similarity score S on +1) m (k) quantifies the alignment of scene contours between the map point cloud and the query point cloud across different sectors. To further improve discriminative ability, this invention uses an overlap loss function to improve this similarity measure:

[0019]

[0020] This overlapping loss strategy implements a weighting mechanism: as shown in equation (10), when feature descriptors do not exist at the same position (all-zero state), no score is assigned; for regions of overlapping activation (where both features exist), positive feedback (s) is applied. a >0); For regions with conflict activation (only one feature activity), apply negative feedback (s b <0). This method achieves a dynamic weighted feature activation mode through equation (9), which effectively enhances the robustness of scene structure feature matching.

[0021]

[0022] The optimal alignment position is determined by equation (11), which determines the sliding window position that generates the maximum similarity score. From the obtained complete window similarity distribution, the peak position k is extracted in this invention. * As the optimal matching result of this invention, this directly corresponds to the initial rotation angle estimation between heterogeneous point clouds. This method effectively solves the dimension mismatch problem, while significantly improving the reliability of cross-sensor matching through fine evaluation of feature activation patterns. After completing the traversal of all map frames, this invention further selects the top Num points with the highest similarity scores. K A map is generated to form a sequence of candidate frames, which serves as input for subsequent processing modules.

[0023] Step S3. The preliminary retrieval module only provides two-dimensional position identification and initial rotation estimation. Projecting the point cloud onto the radial-elevation plane inevitably leads to the loss of three-dimensional structural information—which may result in mismatches. To overcome these limitations, this invention designs a verification module based on voxel space alignment. This module effectively compensates for the limitations of the feature matching stage through precise three-dimensional geometric consistency checks, ensuring comprehensive and accurate attitude estimation when performing reliable localization in complex environments.

[0024] First, this invention performs a process 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, this invention performs GICP based on the initial rotational pose provided during the preliminary retrieval phase. Finally, this invention evaluates the matching quality by calculating the overlap metric between registered point clouds.

[0025]

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

[0027] The beneficial effects of this invention are as follows:

[0028] The present invention demonstrated superior performance in eight sequences of three heterogeneous lidar scenarios tested: the median translation error was generally maintained below 0.5, and the span of the box and the beard was significantly smaller than that of the comparative methods (such as BTC and BEVPlace). This not only reflects the high positioning accuracy of the algorithm, but also shows that its results have good stability and consistency, fully demonstrating the robustness and reliability of the present invention in translation error control. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the implementation of the heterogeneous point cloud global localization method based on a multi-sector loss strategy as described in this invention.

[0030] Figure 2 This is a radar chart comparing the method described in this invention with other open-source methods on various datasets;

[0031] Figure 3 This is a time consumption analysis diagram of each module in the method described in this invention. Detailed Implementation

[0032] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0033] As shown in the figure, the heterogeneous point cloud global localization method based on a multi-sector loss strategy described in this invention comprises the following steps:

[0034] Step 1: The heterogeneous point cloud global localization method based on multi-sector loss strategy as described in claim 1, characterized in that, in step 1, the point cloud from any type of LiDAR is divided into sectors and projected onto the radial-elevation plane of a cylindrical coordinate system to obtain the occupancy feature matrix; specifically as follows:

[0035] When processing the input point cloud Q, this invention eliminates the inherent variations in horizontal and vertical FoV between different LiDAR models through cylindrical coordinate transformation. Unlike the spherical coordinate system used in existing methods, the cylindrical method of this invention provides superior adaptability by maintaining the laser radar beam distribution characteristics with radial resolution through axial decomposition. For each point p in the point cloud Q, initially defined as p = {x, y, z} in Cartesian coordinates, this invention performs a transformation to cylindrical coordinates p = {r, θ, z} according to formula (13):

[0036]

[0037] The method of this invention employs a divide-and-conquer strategy, processing point clouds through sectorization. First, this invention discretizes the point cloud within each sector using voxelization. Then, this invention projects these 3D voxels onto the radial-elevation plane of the corresponding sector, generating an N-layer BEV feature map through hierarchical projection. This representation possesses inherent rotation invariance through polar coordinate decomposition, effectively preserving the scene contour features within each sector.

[0038] Step 2: To further enhance feature compatibility among heterogeneous point clouds, this 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 for solving the feature unification problem of heterogeneous point clouds with different field of view angles.

[0039] Furthermore, to address the dimensionality mismatch issue of feature descriptors from heterogeneous LiDAR systems, this invention proposes an overlap loss matching mechanism based on a sliding window. Consider a map frame feature matrix F constructed from an Ouster-128 LiDAR system. O (Dimension is N) O ×(R·H)), as a reference sequence, and a feature matrix F constructed from a Livox-Mid70 lidar. L (Dimension is N) L ×(R·H)), as the query sequence (where N is N)O >N L This invention achieves heterogeneous point cloud matching by constructing an extended matrix, the matrix of which has a dimension of (N). O +N L -1)×(R·H), define a size of N L Sliding window of ×(R·H) At position k, the basic similarity can initially be formulated as:

[0040]

[0041] Each sliding window position k (1≤k≤N) O -N L The similarity score S on +1) m (k) quantifies the alignment of scene contours between the map point cloud and the query point cloud across different sectors. To further improve discriminative ability, this invention uses an overlap loss function to improve this similarity measure:

[0042]

[0043] This overlapping loss strategy implements a weighted mechanism: as shown in equation (16), when the feature descriptors do not exist at the same position (all-zero state), no score is assigned; for regions of overlapping activation (where both features exist), positive feedback (s) is applied. a >0); For regions with conflict activation (only one feature activity), apply negative feedback (s b <0). This method achieves a dynamic weighted feature activation mode through equation (15), which effectively enhances the robustness of scene structure feature matching.

[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 peak position k is extracted in this invention. * As the optimal matching result of this invention, this directly corresponds to the initial rotation angle estimation between heterogeneous point clouds. This method effectively solves the dimension mismatch problem, while significantly improving the reliability of cross-sensor matching through fine evaluation of feature activation patterns. After completing the traversal of all map frames, this invention further selects the top Num points with the highest similarity scores. K A map is generated to form a sequence of candidate frames, which serves as input for subsequent processing modules.

[0046] Step 3: The preliminary retrieval module only provides two-dimensional position identification and initial rotation estimation. Projecting the point cloud onto the radial-elevation plane inevitably leads to the loss of three-dimensional structural information—which may result in mismatches. To overcome these limitations, this invention has carefully designed a verification module based on voxel space alignment. This module effectively compensates for the limitations of the feature matching stage through precise three-dimensional geometric consistency checks, ensuring comprehensive and accurate attitude estimation when performing reliable localization in complex environments.

[0047] First, this invention performs a process 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, this invention performs GICP based on the initial rotational pose provided during the preliminary retrieval phase. Finally, this invention evaluates the matching quality by calculating the overlap metric between registered point clouds.

[0048]

[0049] Where P and Q represent the voxelized candidate points and query point cloud, respectively, and N is the number of points in the point set P. It is an indicator function, when point p i There is at least one point q within the distance ∈. j When the point set Q is in the function, the function equals 1. When the overlap metric exceeds the threshold Thre... O If the match is valid, the 6-DOF pose optimized by GICP is output; otherwise, a 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 concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A heterogeneous point cloud global localization method based on a multi-sector loss strategy, characterized in that, include: Step 1: Divide the point cloud from any type of LiDAR into sectors and project it onto the radial-elevation plane of the cylindrical coordinate system to obtain the feature matrix. Step 2: Input the flattened and dimensionality-reduced feature matrix into the system, calculate the similarity of sectors between the query frame and the map frame using a sliding window and stacking loss strategy, and generate a candidate queue based on the similarity. To further enhance feature compatibility between heterogeneous point clouds, tensor expansion is used to... The three-dimensional feature matrix is ​​reconstructed as The two-dimensional feature matrix is ​​used. Furthermore, to address the feature matrix dimension mismatch problem, a sliding window-based overlap loss function is proposed. This is based on a map frame feature matrix constructed using the Ouster-128 lidar. Dimension is As a reference sequence, and a query frame feature matrix constructed from a Livox-Mid70 LiDAR. Dimension is , as a query sequence, where Heterogeneous point cloud matching is achieved by constructing an extended matrix, the dimension of which is... Define a size of sliding window In position First, the basic similarity formula without introducing the overlap loss function is formulated as follows: （1）； Each sliding window position Similarity score The alignment degree of scene contours between map point clouds and query point clouds in different sectors was quantified. To further improve the discriminative ability, an overlap loss function was used to improve the basic similarity, resulting in an improved similarity score. （2）； （3）； This overlapping loss function implements a weighting mechanism: as shown in equation (3), no score is assigned when the feature matrix does not exist at the same position; for regions of overlapping activation, positive feedback is applied ( For conflict-activated regions, apply negative feedback ( This method calculates the dynamic weighted feature activation mode through equation (2), which effectively enhances the robustness of scene structure feature matching; （4）； The optimal alignment position is determined by equation (4), which determines the sliding window position that produces the optimal similarity score; from the obtained complete window similarity distribution, the peak position is extracted. As the optimal matching result, this directly corresponds to the initial rotation estimation between heterogeneous point clouds; after completing the traversal of all map frames, the top matching with the highest similarity score is further selected. A map is generated to form a candidate frame sequence, which serves as input for subsequent processing; Step 3: Use the similar position results of sectors in the candidate queue as the initial rotation estimate, and filter out potential erroneous retrievals through the voxel overlap similarity scoring mechanism and GICP technology, thereby outputting accurate 6-DOF pose estimation results.

2. The heterogeneous point cloud global localization method based on a multi-sector loss strategy as described in claim 1, characterized in that, Step 1 involves dividing the point cloud from any type of LiDAR into sectors and projecting it onto the radial-elevation plane of a cylindrical coordinate system to obtain the feature matrix; specifically as follows: For input point cloud Each point in Originally defined as Cartesian coordinates According to formula (5), the transformation to cylindrical coordinates is performed. : （5）； A divide-and-conquer strategy is adopted. First, voxelization is used to discretize the point cloud in each sector. Then, the discretized point cloud is 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 localization method based on a multi-sector loss strategy as described in claim 2, characterized in that, Step 3 is as follows: The initial retrieval module only provides two-dimensional position identification and initial rotation estimation. Projecting the point cloud onto the radial-elevation plane inevitably leads to the loss of three-dimensional structural information. To overcome these limitations, a verification module based on voxel space alignment was designed. The verification module compensates for the limitations of the feature matching stage by checking the three-dimensional geometric consistency, ensuring the comprehensiveness and accuracy of attitude estimation when performing reliable localization in complex environments. First, examine the candidate map point cloud in voxel space. And query point cloud Discretization is performed to standardize the spatial resolution, effectively addressing the challenges posed by heterogeneous point clouds with different resolutions and scanning modes; Subsequently, GICP is performed based on the initial rotation estimate provided in the preliminary retrieval phase; finally, the discretized query point cloud is calculated. Discretized map point cloud The overlap between the two metrics is used to evaluate the quality of the match. （6）； in, It is a query point cloud after discretization The number of points in the middle, It is an indicator function, when the point In distance There is at least one point inside. Discretized map point cloud When the overlap metric exceeds the threshold, the indicator function equals 1; when the overlap metric exceeds the threshold... If the match is valid, the 6-DOF pose optimized by GICP is output; otherwise, a re-retrieval mechanism is triggered to check the next candidate frame in the queue.

Images