Robot pose recognition method based on middle-layer scanning context

Abstract

The invention discloses a robot pose recognition method based on a middle-layer scanning context, and the method comprises the specific steps: 1, point cloud collection and preprocessing: enabling a ship or a water surface robot to operate on the water surface, collecting 3D point cloud data in real time through a laser radar, carrying out the water surface semantic obtaining, inclination angle calculation and coordinate conversion of each frame of original point cloud, and carrying out the point cloud collection and preprocessing; the pitch angle / roll angle error is corrected, the point cloud direction is corrected, and the direction-unified middle-layer point cloud Pmrot is obtained; step 2, sub-descriptor generation and database construction: generating middle-layer scanning context sub-descriptors based on Pmrot, storing all the sub-descriptors and corresponding ship or water surface robot poses, and constructing a position recognition database; and step 3, descriptor similarity matching and real-time position identification: the robot collects a new frame point cloud in real time, and the step 1 is repeated to obtain a new sub descriptor. The rotation invariance of the yaw angle, the pitch angle and the roll angle is realized, and the position identification precision and efficiency of the ship or the water surface robot are improved.

Description

Technical Field

[0001] This invention relates to the field of ship or surface robot positioning and navigation technology. Specifically, it is a method for ship or surface robot position recognition based on 3D point cloud surface semantics in scenarios with pitch and roll angle errors. It can be applied to ship or surface robot simultaneous localization and mapping (SLAM) systems to improve the positioning accuracy and stability of ships or surface robots in purely wave environments. Background Technology

[0002] Simultaneous Localization and Mapping (SLAM) is a core technology for enabling autonomous navigation in various robots. Position recognition, as a key component of SLAM, helps robots correct accumulated sensor errors, identify visited areas, and ensure positioning accuracy. Currently, mainstream position recognition methods are mainly divided into two categories: vision-based and LiDAR-based. 1. Vision-based location recognition methods: These methods often employ the bag-of-words (BOW) model, achieving localization by matching the similarity of real-time acquired images with images in a database. However, this method is susceptible to environmental factors such as changes in lighting and seasonal variations, resulting in poor robustness.

[0003] 2. Location identification method based on lidar: Because lidar is not sensitive to changes in ambient light and weather, it is more robust and widely used in complex scenarios. It is further divided into local positioning and global positioning. Local localization requires an initial position to be known and estimates the local pose based on sensor observation parameters. Although it pursues timeliness, it cannot solve the problem of localization failure caused by accumulated sensor errors. Global localization achieves accurate positioning by efficiently identifying the scenes experienced by the robot and can correct accumulated errors in real time. Global descriptors (such as Scan Context, M2DP, etc.) have become a research hotspot because they do not require a lot of local computation and are highly efficient in extracting global features.

[0004] However, existing LiDAR-based position recognition methods have significant drawbacks: when a robot is running in a non-flat environment (such as water), the undulation of the water surface will cause changes in the robot's pitch and roll angles. However, existing methods only focus on yaw angle errors and ignore the impact of pitch and roll angle errors on the LiDAR scanning range. The LiDAR scanning range shifts under different postures, causing point cloud data at the same location to be unable to be directly matched, ultimately leading to positioning failure.

[0005] For example, while the existing Scan Context method can achieve yaw angle rotation invariance and preserve the internal structure of the point cloud to improve recognition accuracy, it does not consider pitch and roll angle errors. In scenarios with water surface fluctuations, changes in the scanning range can cause misalignment of point cloud features, resulting in a significant decrease in positioning accuracy. The M2DP method extracts global descriptors by projecting 3D point clouds onto multiple 2D planes, but it also fails to solve the attitude error problem in non-flat environments and cannot meet the positioning requirements of ships or robots under water surface conditions.

[0006] Existing related patent technologies, such as the self-moving robot positioning method, device, and self-moving robot disclosed in patent document (CN116989783A), are applicable to scenarios such as "hotels, restaurants, and inspections," which are all indoor flat ground environments, not unstable environments such as water surfaces. Water surface environments cause the robot to continuously generate pitch and roll movements. The technical solution of this patent does not consider the impact of dynamic posture changes on positioning, and the algorithm used is based on the assumption of a relatively stable ground environment. When applied to unstable environments such as water surfaces, the laser point cloud data will show significant differences due to the continuous changes in robot posture, leading to the failure of feature matching and point cloud registration. It has significant limitations, especially in handling three-dimensional posture changes (particularly pitch and roll angles), and is not suitable for application in environments with continuously changing robot postures, such as water surfaces.

[0007] Therefore, there is an urgent need for a position identification method that can simultaneously handle yaw, pitch, and roll angle errors and maintain high positioning accuracy in a water environment. Summary of the Invention

[0008] To address the problems of existing robot position recognition methods neglecting pitch and roll angle errors and easily failing in water environments, this invention provides a robot pose recognition method based on middle layer scan context. By extracting semantics from the water surface, transforming coordinates, filtering middle layer point clouds, and fusing multiple features, the method achieves rotational invariance of yaw, pitch, and roll angles, thereby improving the accuracy and efficiency of position recognition for ships or water surface robots.

[0009] To achieve the above objectives, the technical solution of this invention is: a robot pose recognition method based on mid-level scanning context. Through water surface semantic extraction, coordinate transformation, mid-level point cloud filtering, and multi-feature fusion, it achieves rotational invariance of yaw, pitch, and roll angles, improving the accuracy and efficiency of position recognition for ships or surface robots. Specific steps include: Step 1, Point Cloud Acquisition and Preprocessing: While the ship or surface robot is operating on the water surface, 3D point cloud data is acquired in real time using LiDAR. Water surface semantic acquisition, tilt angle calculation, and coordinate transformation are performed on each frame of the original point cloud. Pitch / roll angle errors are corrected, and point cloud orientation is corrected to obtain a mid-level point cloud with unified orientation. P m_rot Step 2, Sub-descriptor generation and database construction: Based on P m_rot Generate mid-level scanning context sub-descriptors, store all sub-descriptors with their corresponding ship or surface robot poses, and construct a position recognition database; Step 3, descriptor similarity matching, real-time position recognition: The robot acquires new frame point clouds in real time, and repeats step 1 to obtain new sub-descriptors. .

[0010] Furthermore, in step 1, the semantic acquisition of the water surface in each frame of the original point cloud specifically involves: extracting the water surface point cloud from the original lidar point cloud using the Random Sample Consensus (RANSAC) algorithm. The specific process is as follows: Let the original lidar point cloud be... P ={ p 1, p 2,…, p n},in p i ={ x i , y i , z i} represents the coordinates of a single point cloud in the lidar coordinate system; the target extracts a subset of the water surface point cloud. P G Randomly from P Select four points and calculate the parameters of the plane they form. A , B , C , D Satisfying the plane equation Iteratively select different combinations of 4 points, calculate the number of point clouds contained in each plane, and retain the plane with the most point clouds as the water surface plane; the corresponding point cloud is the initial water surface point cloud; introduce a distance threshold. e Calculate the distance from the point cloud to the water surface. Only keep d< e The point cloud as the final water surface point cloud PG This eliminates extraction errors caused by sensor errors and uneven water surfaces.

[0011] Furthermore, in step 1, the tilt angle calculation and coordinate transformation are based on the water surface point cloud. P G Construct a new coordinate system and correct for pitch and roll errors: Calculate the unit normal vector of the water surface plane. , denoted as ( r x , r y , r z Set it as the Z-axis of the new coordinate system; construct the coordinate transformation moments. R Among them, yaw angle i y = 0, by solving the pitch angle i p and roll angle i r Determine the matrix parameters:

[0012] For the original point cloud P Perform coordinate transformation to obtain the transformed point cloud. .

[0013] Furthermore, the tilt angle calculation and coordinate transformation also include verifying the effectiveness of the transformation: calculating... P t The average height, compared to the actual installation height of the lidar. h If the height difference exceeds the error range, repeat step 1 to extract the ground point cloud to ensure the accuracy of coordinate transformation.

[0014] Furthermore, tilt angle calculation and coordinate transformation also include filtering intermediate point clouds. P m Considering the overlapping areas of the lidar scanning range at different tilt angles, the height range of the mid-level point cloud is set as follows: ,in l The scanning radius of the lidar. i To ensure a consistent scanning range of the mid-level point cloud under different tilt angles.

[0015] Furthermore, in step 1, point cloud orientation correction is performed using principal component analysis (PCA) to achieve yaw angle rotation invariance and unify the orientation of point clouds at the same location. Calculate mid-level point clouds P m center coordinates ,in n forP m Number of mid-point clouds Construct the covariance matrix (C):

[0016] Solve the covariance matrix C eigenvalues ​​λ and eigenvectors p satisfy The eigenvector corresponding to the largest eigenvalue is selected as the X-axis of the new coordinate system, the original Z-axis is retained, and the Y-axis is determined by the right-hand rule; the middle-layer point cloud is then... P m Rotate around the Z-axis to unify the orientation of the point cloud at different yaw angles, thus obtaining the orientation-corrected mid-level point cloud. P m_rot .

[0017] Furthermore, in step 2, based on P m_rot The specific steps for generating the mid-level scan context sub-descriptor are as follows: based on the mid-level point cloud after orientation correction P m_rot To generate a global descriptor by fusing height and density features and compensate for information loss during point cloud preprocessing, the process begins with subspace partitioning. Using the LiDAR center as the origin, the point cloud space is divided along the circumferential and radial directions. There are 1 subspace, each denoted as _ . S ij , i For circumferential direction index, j The radial direction index is used; next, eigenvalue calculation is performed for each subspace. S ij Calculate its eigenvalues ​​by weighting. a ij By integrating the height of the highest point within the subspace with the number of point clouds, ,in, k 1 and k 2 is the weighting coefficient, max( z ( p )) represents the maximum height of the point cloud within the subspace. Z coordinate, count ( p The number of point clouds in the subspace is represented by ; finally, a descriptor is constructed, which represents the feature values ​​of all subspaces. a ij constitute The two-dimensional matrix is ​​constructed as follows: First, an intermediate layer descriptor based on height information is built; second, an intermediate layer descriptor based on density information is built; finally, an intermediate layer descriptor that fuses height and density information is generated from the two descriptors, which is the intermediate layer scan context descriptor, denoted as . I .

[0018] Furthermore, in step 3, descriptor similarity matching performs a two-stage matching process: filtering by circular key... k 1 candidate descriptor, calculate The candidate descriptor with the minimum cosine distance to the candidate descriptor is selected. If the minimum cosine distance is less than the corresponding threshold, the recognition is successful, and the pose corresponding to the candidate descriptor is output as the current position of the ship or surface robot. Otherwise, it is determined that the visited area has not been recognized, and the database continues to be built.

[0019] Furthermore, the two-stage matching process for descriptor similarity matching specifically includes: employing a two-stage matching algorithm to achieve efficient and accurate descriptor matching, reducing computational costs and ensuring positioning accuracy: The first stage: generating circular keys to filter candidate descriptors, and then... I Each row is converted into a real number (representing a circular feature) through an encoding function, constructing... The circular key; in the database, nearest neighbor search is performed using the circular key to filter out the most similar items. m The first stage involves selecting candidate descriptors to narrow down the matching range and reduce computational cost; the second stage involves performing cosine distance exact matching on the query descriptors. Compared with the filtered m candidate descriptors Calculate the cosine distance between the two to measure their similarity.

[0020] in, , These are the query sub-descriptors and candidate descriptors, respectively. j Column vectors, cosine distance d The smaller the value, the higher the similarity; then the matching result is determined.

[0021] Furthermore, if the minimum cosine distance between the query sub-descriptor and the candidate descriptor is less than a set threshold, and the corresponding actual pose distance is less than the threshold, then the position recognition is considered successful, and the position corresponding to the candidate descriptor is output as the current position of the ship or surface robot.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Robustness to multi-pose errors: By extracting semantics from the water surface and transforming coordinates, the influence of pitch and roll angle errors on the scanning range is eliminated; combined with PCA to achieve rotation invariance of yaw angle, stable positioning of ships or water surface robots is achieved for the first time in scenarios where there are errors in yaw angle, pitch angle and roll angle, solving the problem of positioning failure in the water surface environment; 2. High localization accuracy: The point cloud height and density features are fused during the descriptor generation process to make up for the information loss in the preprocessing stage. Compared with existing methods that rely on only a single feature (such as Scan Context), the localization accuracy is improved by 15%-25% (based on KITTI dataset and self-collected dataset verification). 3. Low computational cost: A two-stage matching algorithm is adopted, which filters candidate sub-descriptors through circular keys to avoid matching the entire database; at the same time, since PCA has implemented yaw angle correction, there is no need to enumerate the column offsets of candidate descriptors. The total computation time is only 2 / 3 of the existing Scan Context method, which meets the requirements of real-time positioning. Attached Figure Description Figure 1 This is the original point cloud image; Figure 2 This is a point cloud map of the intermediate layer; Figure 3 This is a descriptor graph for the intermediate layer based on height information; Figure 4 This is a density-based intermediate layer descriptor graph; Figure 5 An intermediate layer descriptor graph that integrates height and density information. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0024] like Figure 1 to Figure 5 As shown, the robot pose recognition method based on middle layer scan context of the present invention includes point cloud acquisition and preprocessing, middle layer scan context descriptor generation, and descriptor similarity matching. The specific steps are as follows: 1. Point cloud acquisition and preprocessing Point cloud acquisition: Ships or surface robots operate on the water surface and acquire 3D point cloud data in real time through LiDAR; point cloud preprocessing aims to eliminate the influence of pitch and roll angle errors on the scanning range, while achieving yaw angle rotation invariance, including three sub-steps: water surface semantic acquisition, tilt angle calculation and coordinate transformation, and point cloud direction correction.

[0025] 1.1 Semantic Acquisition of Water Surface The Random Sample Consensus (RANSAC) algorithm is used to extract water surface point clouds from the original lidar point cloud. The specific process is as follows: Let the original lidar point cloud be P ={ p 1, p 2,…, p n},in pi ={ x i , y i , z i} represents the coordinates of a single point cloud in the lidar coordinate system; the target extracts a subset of the water surface point cloud. P G Randomly from P Select four points and calculate the parameters of the plane they form. A , B , C , D Satisfying the plane equation Iteratively select different combinations of 4 points, calculate the number of point clouds contained in each plane, and retain the plane with the most point clouds as the water surface plane; the corresponding point cloud is the initial water surface point cloud; introduce a distance threshold. e (Based on the actual scene settings, such as 0.1m), calculate the distance from the point cloud to the water surface. Only keep d< e The point cloud as the final water surface point cloud P G This eliminates extraction errors caused by sensor errors and uneven water surfaces.

[0026] 1.2 Calculation of Inclination Angle and Coordinate Transformation Based on water surface point clouds P G Construct a new coordinate system and correct for pitch and roll errors: Calculate the unit normal vector of the water surface plane. , denoted as ( r x , r y , r z Set it as the Z-axis of the new coordinate system; construct the coordinate transformation moments. R Among them, yaw angle i y = 0, by solving the pitch angle i p and roll angle i r Determine the matrix parameters:

[0027] For the original point cloud P Perform coordinate transformation to obtain the transformed point cloud. .

[0028] Verify the validity of the conversion: Calculate P t The average height, compared to the actual installation height of the lidar.h If the height difference exceeds the error range, repeat step 1.1 to extract the ground point cloud to ensure the accuracy of coordinate transformation.

[0029] Filtering mid-level point clouds P m Considering the overlapping areas of the lidar scanning range at different tilt angles, the height range of the mid-level point cloud is set as follows: ,in l The scanning radius of the lidar. i To ensure a consistent scanning range of the mid-level point cloud under different tilt angles.

[0030] 1.3 Point Cloud Orientation Correction Principal component analysis (PCA) is used to achieve yaw angle rotation invariance, unifying the orientation of point clouds at the same location: Calculate mid-level point clouds P m center coordinates ,in n for P m Number of midpoint clouds.

[0031] Construct the covariance matrix (C):

[0032] Solve the covariance matrix C eigenvalues ​​λ and eigenvectors p (satisfy The eigenvector corresponding to the largest eigenvalue is selected as the X-axis of the new coordinate system, the original Z-axis is retained, and the Y-axis is determined by the right-hand rule; the middle-layer point cloud is then... P m Rotate around the Z-axis to unify the orientation of the point cloud at different yaw angles, thus obtaining the orientation-corrected mid-level point cloud. P m_rot .

[0033] 2. Generation of mid-level scan context descriptors Mid-level point cloud based on orientation correction P m_rot It integrates height and density features to generate a global descriptor, compensating for information loss during point cloud preprocessing.

[0034] First, the subspace is divided. Taking the center of the LiDAR as the origin, the point cloud space is divided along the circumferential direction (angular dimension) and the radial direction (distance dimension). There are 1 subspace, each denoted as _ . S ij , i For circumferential direction index, jThis is the radial direction index.

[0035] Next, eigenvalues ​​are calculated for each subspace. S ij Calculate its eigenvalues ​​by weighting. a ij By integrating the height of the highest point within the subspace with the number (density) of the point cloud, ,in, k 1 and k 2 is the weighting coefficient, max( z ( p )) represents the maximum value of the point cloud within the subspace. Z Coordinates (height) count ( p ) represents the number (density) of point clouds within the subspace; Finally, descriptor construction is performed, taking the feature values ​​of all subspaces. a ij constitute A two-dimensional matrix. First, an intermediate layer descriptor based on height information is constructed. Second, an intermediate layer descriptor based on density information is constructed. Finally, using the two descriptors, an intermediate layer descriptor that fuses the height and density information is generated, which is the intermediate layer scan context descriptor, denoted as . I .

[0036] 3. Descriptor Similarity Matching A two-stage matching algorithm is used to achieve efficient and accurate descriptor matching, reducing computational costs and ensuring positioning accuracy. Phase 1: Generate ring key filtering candidate descriptors. I Each row is converted into a real number (representing a circular feature) through an encoding function, constructing... The circular key; in the database, nearest neighbor search is performed using the circular key to filter out the most similar items. m By using multiple candidate descriptors, the matching range is narrowed down, reducing the computational load.

[0037] Phase Two: Perform cosine distance exact matching on the query descriptor. Compared with the filtered m candidate descriptors Calculate the cosine distance between the two to measure their similarity. .

[0038] in, , These are the query sub-descriptors and candidate descriptors, respectively. j Column vectors, cosine distance dThe smaller the value, the higher the similarity. Then, the matching result is judged. If the minimum cosine distance between the query sub-descriptor and the candidate descriptor is less than the set threshold, and the corresponding actual pose distance (water surface truth value) is less than the threshold, the position recognition is determined to be successful, and the position corresponding to the candidate descriptor is output as the current position of the ship or water surface robot.

Claims

1. A robot pose recognition method based on mid-level scan context, characterized in that, By extracting surface semantics, transforming coordinates, filtering mid-level point clouds, and fusing multiple features, rotational invariance of yaw, pitch, and roll angles is achieved, improving the accuracy and efficiency of position recognition for ships or surface robots. Specific steps include: Step 1, Point Cloud Acquisition and Preprocessing: While the ship or surface robot is operating on the water surface, 3D point cloud data is acquired in real-time using LiDAR. For each frame of the original point cloud, surface semantics are acquired, tilt angles are calculated and coordinates are transformed, pitch / roll angle errors are corrected, and point cloud orientation is corrected to obtain a mid-level point cloud with uniform orientation. P m_rot Step 2, Sub-descriptor generation and database construction: Based on P m_rot Generate mid-level scanning context sub-descriptors, store all sub-descriptors with their corresponding ship or surface robot poses, and construct a position recognition database; Step 3, descriptor similarity matching, real-time position recognition: The robot acquires new frame point clouds in real time, and repeats step 1 to obtain new sub-descriptors. .

2. The robot pose recognition method based on mid-level scanning context according to claim 1, characterized in that, In step 1, the semantic acquisition of the water surface in each frame of the original point cloud specifically involves: extracting the water surface point cloud from the original lidar point cloud using the Random Sample Consensus (RANSAC) algorithm. The specific process is as follows: Let the original lidar point cloud be... P ={ p 1, p 2, … , p n },in p i ={ x i , y i , z i } represents the coordinates of a single point cloud in the lidar coordinate system; the target extracts a subset of the water surface point cloud. P G Randomly from P Select four points and calculate the parameters of the plane they form. A , B , C , D Satisfying the plane equation Iteratively select different combinations of 4 points, calculate the number of point clouds contained in each plane, and retain the plane with the most point clouds as the water surface plane; the corresponding point cloud is the initial water surface point cloud; introduce a distance threshold. ε Calculate the distance from the point cloud to the water surface. Only keep d < ε The point cloud as the final water surface point cloud P G This eliminates extraction errors caused by sensor errors and uneven water surfaces.

3. The robot pose recognition method based on mid-level scanning context according to claim 1, characterized in that, In step 1, the tilt angle is calculated and coordinate transformation is performed based on the water surface point cloud. P G Construct a new coordinate system and correct for pitch and roll errors: Calculate the unit normal vector of the water surface plane. , denoted as ( r x , r y , r z Set it as the Z-axis of the new coordinate system; construct the coordinate transformation moments. R Among them, yaw angle θ y = 0, by solving the pitch angle θ p and roll angle θ r Determine the matrix parameters: For the original point cloud P Perform coordinate transformation to obtain the transformed point cloud. .

4. The robot pose recognition method based on mid-level scan context according to claim 3, characterized in that, Tilt angle calculation and coordinate transformation also include verifying the effectiveness of the transformation: calculation P t The average height, compared to the actual installation height of the lidar. h If the height difference exceeds the error range, repeat step 1 to extract the ground point cloud to ensure the accuracy of coordinate transformation.

5. The robot pose recognition method based on mid-level scanning context according to claim 3, characterized in that, Tilt angle calculation and coordinate transformation also include filtering intermediate point clouds. P m Considering the overlapping areas of the lidar scanning range at different tilt angles, the height range of the mid-level point cloud is set as follows: ,in l The scanning radius of the lidar. θ To ensure a consistent scanning range of the mid-level point cloud under different tilt angles.

6. The robot pose recognition method based on mid-level scan context according to claim 1, characterized in that, In step 1, point cloud orientation correction is performed using principal component analysis (PCA) to achieve yaw angle rotation invariance and unify the orientation of point clouds at the same location. Calculate mid-level point clouds P m center coordinates ,in n for P m Number of mid-point clouds Construct the covariance matrix (C): Solve the covariance matrix C eigenvalues ​​λ and eigenvectors p satisfy The eigenvector corresponding to the largest eigenvalue is selected as the X-axis of the new coordinate system, the original Z-axis is retained, and the Y-axis is determined by the right-hand rule; the middle-layer point cloud is then... P m Rotate around the Z-axis to unify the orientation of the point cloud at different yaw angles, thus obtaining the orientation-corrected mid-level point cloud. P m_rot .

7. The robot pose recognition method based on mid-level scan context according to claim 1, characterized in that, In step 2, based on P m_rot The specific steps for generating the mid-level scan context sub-descriptor are as follows: based on the mid-level point cloud after orientation correction P m_rot To generate a global descriptor by fusing height and density features and compensate for information loss during point cloud preprocessing, the process begins with subspace partitioning. Using the LiDAR center as the origin, the point cloud space is divided along the circumferential and radial directions. There are 1 subspace, each denoted as _ . S ij , i For circumferential direction index, j The radial direction index is used; next, eigenvalue calculation is performed for each subspace. S ij Calculate its eigenvalues ​​by weighting. a ij By integrating the height of the highest point within the subspace with the number of point clouds, ,in, k 1 and k 2 is the weighting coefficient, max( z ( p )) represents the maximum height of the point cloud within the subspace. Z coordinate, count ( p The number of point clouds in the subspace is represented by ; finally, a descriptor is constructed, which represents the feature values ​​of all subspaces. a ij constitute The two-dimensional matrix is ​​constructed as follows: First, an intermediate layer descriptor based on height information is built; second, an intermediate layer descriptor based on density information is built; finally, an intermediate layer descriptor that fuses height and density information is generated from the two descriptors, which is the intermediate layer scan context descriptor, denoted as . I .

8. The robot pose recognition method based on mid-level scan context according to claim 1, characterized in that, In step 3, descriptor similarity matching performs a two-stage matching process: filtering by circular key... k 1 candidate descriptor, calculate The candidate descriptor with the minimum cosine distance to the candidate descriptor is selected. If the minimum cosine distance is less than the corresponding threshold, the recognition is successful, and the pose corresponding to the candidate descriptor is output as the current position of the ship or surface robot. Otherwise, it is determined that the visited area has not been recognized, and the database continues to be built.

9. The robot pose recognition method based on mid-level scan context according to claim 8, characterized in that, Descriptor similarity matching performs two-stage matching, specifically including: employing a two-stage matching algorithm to achieve efficient and accurate descriptor matching, reducing computational costs and ensuring positioning accuracy: The first stage: generating circular keys to filter candidate descriptors, and then... I Each row is converted into a real number through an encoding function, representing a circular feature, thus constructing... The circular key; in the database, nearest neighbor search is performed using the circular key to filter out the most similar items. m The first stage involves selecting candidate descriptors to narrow down the matching range and reduce computational cost; the second stage involves performing cosine distance exact matching on the query descriptors. Compared with the filtered m candidate descriptors Calculate the cosine distance between the two to measure their similarity. in, , These are the query sub-descriptors and candidate descriptors, respectively. j Column vectors, cosine distance d The smaller the value, the higher the similarity; then the matching result is determined.

10. The robot pose recognition method based on mid-level scan context according to claim 9, characterized in that, If the minimum cosine distance between the query sub-descriptor and the candidate descriptor is less than a set threshold, and the corresponding actual pose distance is less than the threshold, then the position recognition is successful, and the position corresponding to the candidate descriptor is output as the current position of the ship or surface robot.

Images