Memory self-updating curbstone storage and control method and device and storage medium

Abstract

The invention discloses a memory self-updating curbstone storage and control method and device and a storage medium. The method comprises the following steps: S1, feature preprocessing and confidence coefficient design: carrying out 3D information recovery on multi-path curbstone detection information, then carrying out subgraph construction and smoothing, and carrying out confidence coefficient design according to the confidence coefficient of a detection result and a subgraph common-view result; s2, splicing and updating the sub-map and the memory map: if the memory map does not exist, generating the memory map by using the result of the sub-map in the step S1; if the memory map exists, splicing is performed by performing nearest neighbor matching and minimum re-projection error optimization on the sub-maps and the memory map, a search tree is constructed for the spliced memory map, a common-view area of a detection result and the memory map is searched, confidence judgment logic is triggered, and the memory map is updated; and S3, post-processing and slicing and tiling the memory map: smoothing the updated memory map to remove false detection points, performing tiled slicing operation on a smoothed result, and storing the result in data in a block form.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and in particular to a memory self-updating curb storage and control method, device, and storage medium. Background Art

[0002] With the development of autonomous driving and intelligent transportation systems, accurate perception and rapid update of the road environment have become increasingly important. As an important feature of the road boundary, the curb is of great significance for vehicle navigation and path planning, especially in the scenario of vehicle edge operation.

[0003] Currently, image perception has the common problem of poor recognition effect in scenarios such as poor light and heavy rain, and there will occasionally be cases of missed detection and false detection. Therefore, it is necessary to construct a memory map and use it as prior information for real-time detection.

[0004] Traditional methods of pre-constructing high-precision maps have problems such as high mapping cost, long update cycle, data privacy and security. Therefore, it is necessary for the vehicle to have the function of quickly constructing a memory map and real-time updating. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a memory self-updating curb storage and control method, device, and storage medium. A memory self-updating curb storage and control method includes the following steps: S1 Feature preprocessing and confidence design: Restore the 3D information of multi-path curb detection information, then construct and smooth submaps, and then perform confidence design according to the confidence of the detection result itself and the co-visibility result of the submaps; S2 Submap and memory map splicing and updating: If there is no memory map, use the result of the submap in step S1 to generate a memory map; if there is a memory map, perform splicing by nearest neighbor matching and minimum reprojection error optimization on the submap and the memory map, construct a search tree for the spliced memory map, find the co-visible area of the detection result and the memory map, trigger the confidence determination logic, and update the memory map; S3 Memory map post-processing and tiling: Perform a smoothing operation on the updated memory map to remove false detection points, and perform a tiling operation on the smoothed result to store it in the data in the form of blocks.

[0006] Further, step S1 specifically includes: S11: Obtain visual detection information and laser 3D detection information, and perform depth restoration on the visual detection information to obtain visual 3D detection information; S12: Perform time synchronization operations through a separate time-triggered thread. When receiving visual 3D detection information and laser 3D detection information in different threads, use the SLAM pose for time synchronization. The formula is as follows: where X target is the vehicle body coordinate value in the target time, X sourceis the coordinate value of the vehicle body coordinate system at the source time, T target is the 4x4 transformation matrix of the SLAM pose at the target time, T source is the 4x4 transformation matrix of the SLAM pose at the source time; S13: Perform Ransac fitting on the synchronized visual 3D detection information and laser 3D detection information; S14: Use the SLAM pose and the result of step S13 to construct a rough submap within the range of N meters and the maximum number of point clouds; S15: Perform MeanShift and downsampling operations on the rough submap to obtain a smooth submap; S16: Perform nearest neighbor search on the smooth submap results by category, and design the confidence according to its own confidence and the co-visible area.

[0007] Further, in step S16, when performing confidence design, for the co-visible area, that is, when a certain area is observed by multiple detectors or viewpoints, and the confidence of the detection results of at least one category is relatively high, the overall confidence of the co-visible area is also relatively high; for the non-co-visible area, that is, the area observed by only one detector or viewpoint, directly use the confidence provided by each detector as the confidence evaluation of this area.

[0008] Further, in step S11, when performing depth recovery, a laser-assisted depth recovery method or an inverse perspective mapping IMP method is adopted.

[0009] Further, step S2 specifically includes: S21: Perform projection transformation on the submap in step S15 using the SLAM pose to generate a submap in the global coordinate system, that is, a global submap, and determine whether there is a memory map at this time. If not, save the global submap at this time as the memory map, otherwise execute the following steps S22 - S29: S22: Read the global submap features and construct a binary search tree of Kd-tree point clouds for storage; S23: Make Euclidean distance nearest neighbor matching between the memory map feature points and the global feature points by category as the candidate matching point pair set, and at the same time perform certain threshold screening, Ransac fitting and outlier filtering operations to obtain the nearest matching point pair set; S24: Set the maximum number of iterations and the iteration termination condition, and the initial pose of each iteration is the pose optimized in the previous iteration; S25: Use all the matching point pairs and poses to construct the following residuals: where R and t are optimization variables, P lidar_i and P map_iare the global sub - graph feature points and the memory map feature points of the i - th matching point pair, and n represents the number of matching point pairs; S26: Use Levenberg–Marquardt to optimize the pose and minimize the residual e in step S25; S27: Repeat steps S23 - S25 until the iteration condition is met, and output the optimized pose; S28: Perform coordinate transformation on the memory map features through the optimized pose and splice them with the global sub - graph; S29: Perform a confidence logic judgment on the spliced memory map. If there is a co - visibility area between the global sub - graph and the initial memory map, use the one with the larger confidence of the two as the updated content; if there is no co - visibility area, use the spliced result.

[0010] Further, in step S1, the multi - path curb detection information includes panoramic camera detection information, fisheye camera detection information, and laser 3D detection information.

[0011] Further, in step S3, by performing clustering and down - sampling operations on the updated memory map, the operations of smoothing and noise reduction of points are completed.

[0012] The technical solution of the present invention also provides a memory self - updating curb storage and control device, which includes the following modules: a feature pre - processing and confidence design module, which is used to restore 3D information from the multi - path curb detection information, then construct and smooth the sub - graph, and perform confidence design according to the confidence of the detection result itself and the co - visibility result of the sub - graph; a sub - graph and memory map splicing and updating module, which is used to generate a memory map using the result of the sub - graph in the feature pre - processing and confidence design module when there is no memory map; when there is a memory map, perform splicing by nearest neighbor matching and minimum reprojection error optimization on the sub - graph and the memory map, construct a search tree for the spliced memory map, find the co - visibility area between the detection result and the memory map, trigger the confidence judgment logic, and update the memory map; a memory map post - processing and tiling module, which is used to perform a smoothing operation on the updated memory map to remove mis - detected points, and perform tiling and cutting operations on the smoothed result, and store it in the data in the form of blocks.

[0013] The technical solution of the present invention also provides a computer - readable storage medium containing a computer program, which, when executed by one or more processors, executes the memory self - updating curb storage and control method as described above.

[0014] The technical solution of the present invention has the following beneficial effects:

[0015] (1) It can receive multi - path curb detection results such as laser detection and image detection; and can perform 2D - to - 3D depth restoration on the image detection results;

[0016] (2) It can perform preprocessing operations such as time synchronization, filtering, fitting, and smoothing on the multi-path detection results, and the results are more robust;

[0017] (3) Subgraph superposition is performed on the above results through the SLAM pose, and the confidence design of the curb points is carried out according to the detection results and the subgraph results, and the subgraph confidence attribute is given;

[0018] (4) The fused curb in the memory map is searched through the SLAM pose, and the detection results are matched and optimized with the memory map to complete the splicing of the detection and the memory map;

[0019] (5) A memory map search tree is constructed to find the common observation area of the detection results and the memory map, trigger the confidence determination logic, and perform an update operation on the memory map;

[0020] (6) Post-processing operations such as clustering and fitting are performed on the updated memory map; and after being cut into tiles and put into the database, the memory occupancy is reduced and the user can obtain the curb information around the vehicle body without feeling it. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 is the system block diagram of the memory self-updating curb storage and control device of the present invention;

[0023] Figure 2 is the flowchart of the feature preprocessing and confidence module of the present invention;

[0024] Figure 3 is the flowchart of the subgraph and memory map splicing and updating of the present invention;

[0025] Figure 4 is the superimposed display diagram of the real-time detection results and the memory map, where the line d3 is the subgraph result generated after real-time detection, and the ones starting with m are the results in the previous memory map;

[0026] Figure 5 is the schematic diagram of the result of the memory map update of the present invention. The white line in the middle of the image is the automatically updated memory map, and the others are the previously existing memory maps;

[0027] Figure 6 is the example diagram of the finally updated memory map of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] As Figure 1 shown, the input of the present invention is the multi-path curb detection of a camera or a radar, and the pose of SLAM (Simultaneous Localization and Mapping); the output is a memory tile map that can be used by a planning or other module.

[0030] The present invention consists of three parts: feature preprocessing and confidence design, stitching and updating of submaps and memory maps, and postprocessing and tiling of memory maps.

[0031] S1 Feature preprocessing and confidence design: Perform 3D information restoration on the multi-path curb detection, and then complete the construction and smoothing of the submap through operations such as time synchronization, Ransac fitting, MeanShift smoothing, and outlier filtering. Then, perform confidence design based on the confidence of the detection itself and the co-visibility results of the submaps.

[0032] S2 Stitching and updating of submaps and memory maps: If there is no memory map, use the results of the above submaps to generate a memory map; if there is a memory map, perform stitching by performing nearest neighbor matching and minimum reprojection error optimization on the submaps and the memory map; construct a search tree for the stitched memory map, find the co-visibility area between the detection results and the memory map, trigger the confidence determination logic, and update the memory map.

[0033] S3 Postprocessing and tiling of memory maps: Perform smoothing operations such as clustering and fitting on the updated memory map to remove misdetection points, and perform tiling operations on the smoothed results to store them in the data in the form of blocks for convenient access.

[0034] The following elaborates on each part in detail.

[0035] As Figure 2 shown, the specific steps of the S1 feature preprocessing and confidence module are as follows:

[0036] S11. Obtain one or more of visual information and laser 3D information; if there is a visual curb detection result, depth recovery of the visual curb is required. In practical applications, the depth recovery methods include: laser-assisted depth recovery, that is, depth information is assigned to the projection points after the laser is projected onto the image. This method requires laser information on the vehicle and has high precision; the IPM (Inverse Perspective Mapping) method that assumes the curb ground is flat to obtain its depth information. This method is faster, but its precision and stability are lower than the first one. Finally, obtain the 3D information of the detection result.

[0037] S12. Perform time synchronization operations in a separate time-triggered thread; when different threads receive visual 3D information and laser 3D information, perform time synchronization on them using the SLAM pose. The formula is as follows:

[0038]

[0039] where, X target is the coordinate value of the vehicle body coordinate system at target time, X source is the coordinate value of the vehicle body coordinate system at source time, T target is the 4×4 transformation matrix of the SLAM pose at target time (i.e., from the vehicle body coordinate system to the slam coordinate system), T source is the 4×4 transformation matrix of the SLAM pose at source time.

[0040] S13. Perform Ransac fitting on the synchronized 3D information to reduce the influence of misdetection noise points;

[0041] S14. Use the SLAM pose and the result of S13 to construct a rough submap within a range of N meters (such as set to 200 - 500m) and the maximum number of point clouds;

[0042] S15. Perform operations such as MeanShift and downsampling on the above rough submap again to obtain a smooth submap; this operation can make the submap result smoother, reduce outliers, and ensure the robustness of subsequent stitching;

[0043] S16. Perform nearest neighbor search on the above results by category (here the category refers to the detection results of different sensors), and design the confidence according to its own confidence and the common view area; if the common view area and at least one category of detection results have high confidence, then the overall confidence is relatively high. In the non-common view area, select their respective confidences. The "relatively high" here means when it reaches the trustworthy value of the detection result; for example, a point confidence of 0.8 can be set here.

[0044] The above confidence level usually represents the degree of certainty of the model regarding its prediction results (such as a certain object category), which is a numerical value between 0 and 1. The higher the value, the more confident the model is in its prediction results. The co-visible area refers to the image area that can be observed simultaneously by multiple detectors or perspectives.

[0045] Based on these two concepts, the present invention can design the following confidence level evaluation strategy:

[0046] (1) Processing of the co-visible area: When a certain area is observed simultaneously by multiple detectors or perspectives (i.e., the co-visible area), we will check whether the detection results of at least one method (such as laser or vision) within this area have a high confidence level. If this condition is met, that is, the confidence levels of the detection results of at least one category are relatively high, then we can consider that the overall confidence level of this co-visible area is also high. This is because the mutual confirmation of multiple detectors increases the reliability of the results.

[0047] (2) Processing of non-co-visible areas: For those areas that are only observed by one detector or perspective (i.e., non-co-visible areas), we cannot use the confirmation of co-visibility to enhance the confidence level. Therefore, for these areas, we will directly adopt the confidence levels provided by their respective detectors as the confidence level evaluation of this area.

[0048] Through such a design, the present invention can more flexibly handle the confidence level issues of different areas and different detection results. In the co-visible area, through the mutual verification of multiple detectors, we can improve the accuracy and robustness of the overall system; while in the non-co-visible area, it depends on the performance of a single detector. This strategy not only takes into account the complementary advantages of multiple perspectives but also retains the independence of single-perspective detection, which helps to improve the comprehensive performance of the entire system.

[0049] As Figure 3 shown, in S2, if there is no memory map, the results of the above submaps are used to generate a memory map; if there is a memory map, then the above submaps and the memory map are subjected to matching and optimization splicing, and then the memory map is updated. The specific steps are as follows:

[0050] S21, perform a projection transformation on the S16 submap using the SLAM pose to generate a submap in the global coordinate system; determine whether there is a memory map at this time. If not, save the global submap at this time as the memory map; if there is a memory map, perform the following steps:

[0051] S22, read the global submap features and construct a binary search tree of the Kd-tree point cloud for storage;

[0052] S23. Classify the memory map feature points and the global feature points by category and perform the nearest neighbor matching of Euclidean distance as the candidate matching point pair set. At the same time, perform certain threshold screening, Ransac fitting, and outlier filtering operations to obtain the nearest matching point pair set;

[0053] S24. Set the maximum number of iterations and the iteration termination condition. The initial pose T for each iteration is the pose optimized in the previous iteration;

[0054] S25. Use all the matching point pairs Match{plidar (global submap), pmap (memory map)’} and the pose Pose to construct the following residuals:

[0055]

[0056] where R and t are optimization variables, P lidar_i and P map_i are the global submap feature point and the memory map feature point of the i-th matching point pair, and n represents the number of matching point pairs;

[0057] S26. Use Levenberg–Marquardt to optimize Pose and minimize the residual e in the above steps;

[0058] S27. Continuously repeat steps S23, S24, and S25 until the iteration condition is met, and output the optimized pose;

[0059] S28. After steps S22 - S27 are completed, perform coordinate transformation on the memory map features through the optimized Pose and splice them with the global submap

[0060] S29. Perform a confidence logic judgment on the spliced memory map: If there is a co-visible area between the global submap and the initial memory map, use the one with the larger confidence of the two as the updated content; if there is no co-visibility, use the spliced result.

[0061] Next, perform clustering and downsampling operations on the above-mentioned memory map to complete the operations of smoothing and noise reduction of points. Finally, through tiling operations, cut it into block maps and store them in the database for subsequent modules to read quickly.

[0062] In an embodiment of the present invention, a method for memory self-update curb storage and control is provided, which includes the following steps: S1 Feature preprocessing and confidence design: restoring 3D information from multi-path curb detection information, then constructing and smoothing subgraphs, and then performing confidence design based on the confidence of the detection results themselves and the co-visibility results of the subgraphs; S2 Subgraph and memory map stitching and update: If there is no memory map, use the results of the subgraphs in step S1 to generate a memory map; if there is a memory map, perform stitching by performing nearest neighbor matching and minimum reprojection error optimization on the subgraphs and the memory map, construct a search tree for the stitched memory map, find the co-visibility area between the detection results and the memory map, trigger the confidence determination logic, and update the memory map; S3 Memory map post-processing and tiling: Perform a smoothing operation on the updated memory map to remove misdetection points, and perform a tiling operation on the smoothed result to store it in the data in the form of blocks.

[0063] Further, step S1 specifically includes: S11: Obtain visual detection information and laser 3D detection information, and perform depth restoration on the visual detection information to obtain visual 3D detection information; S12: Perform time synchronization operations through a separate time-triggered thread. When the visual 3D detection information and the laser 3D detection information are received in different threads, use the SLAM pose for time synchronization. The formula is as follows: where, X target is the vehicle body coordinate value at the target time, X source is the vehicle body coordinate value at the source time, T target is the 4*4 transformation matrix of the SLAM pose at the target time, T source is the 4*4 transformation matrix of the SLAM pose at the source time; S13: Perform Ransac fitting on the synchronized visual 3D detection information and laser 3D detection information; S14: Use the SLAM pose and the results of step S13 to construct a rough subgraph within an N-meter range and a maximum number of point clouds; S15: Perform MeanShift and downsampling operations on the rough subgraph to obtain a smoothed subgraph; S16: Perform nearest neighbor search on the results of the smoothed subgraphs for different categories, and perform confidence design based on their own confidence and co-visibility areas.

[0064] Further, in step S16, when performing confidence design, for the co-visibility area, that is, when a certain area is observed by multiple detectors or viewpoints, when the confidence of the detection results of at least one category is relatively high, the overall confidence of the co-visibility area is also relatively high; for the non-co-visibility area, that is, the area observed by only one detector or viewpoint, directly use the confidence provided by each detector as the confidence evaluation of this area.

[0065] Further, in step S11, when performing depth recovery, a laser-assisted depth recovery method or an inverse perspective mapping (IMP) method is adopted.

[0066] Further, step S2 specifically includes: S21: Projectively transform the sub-graph in step S15 using the SLAM pose to generate a sub-graph in the global coordinate system, i.e., a global sub-graph, and determine whether there is a map in memory at this time. If not, save the current global sub-graph as the map in memory; otherwise, execute the following steps S22-S29: S22: Read the global sub-graph features and construct a binary search tree of the Kd-tree point cloud for storage; S23: Classify the feature points of the map in memory and the global feature points and perform the nearest neighbor matching of the Euclidean distance as a set of candidate matching point pairs. At the same time, perform certain threshold screening, Ransac fitting, and outlier filtering operations to obtain the set of nearest matching point pairs; S24: Set the maximum number of iterations and the iteration termination condition. The initial pose of each iteration is the pose optimized in the previous iteration; S25: Use all the matching point pairs and poses to construct the following residual: where R and t are optimization variables, P lidar_i and P map_i are the global sub-graph feature points and the map in memory feature points of the i-th matching point pair, and n represents the number of matching point pairs; S26: Use the Levenberg–Marquardt method to optimize the pose and minimize the residual e in step S25; S27: Repeat steps S23-S25 until the iteration condition is met, and output the optimized pose; S28: Perform coordinate transformation on the map in memory feature through the optimized pose and splice it with the global sub-graph; S29: Perform a confidence logic judgment on the spliced map in memory. If there is a co-visible area between the global sub-graph and the initial map in memory, use the one with the larger confidence of the two as the updated content; if there is no co-visible area, use the spliced result.

[0067] Further, in step S1, the multi-channel curb detection information includes panoramic camera detection information, fisheye camera detection information, and laser 3D detection information.

[0068] Further, in step S3, by performing clustering and downsampling operations on the updated map in memory, the operations of smoothing and noise reduction of points are completed.

[0069] In another embodiment of the present invention, a memory self-updating curb storage and control device is further provided, which includes the following modules: a feature preprocessing and confidence design module, configured to restore 3D information of multi-path curb detection information, then construct and smooth submaps, and perform confidence design according to the confidence of the detection result itself and the co-visibility result of the submaps; a submap and memory map splicing and updating module, configured to generate a memory map using the result of the submap in the feature preprocessing and confidence design module when there is no memory map; when there is a memory map, perform splicing by performing nearest neighbor matching and minimum reprojection error optimization on the submap and the memory map, construct a search tree for the spliced memory map, find the co-visibility area of the detection result and the memory map, trigger the confidence determination logic, and update the memory map; a memory map post-processing and tiling module, configured to perform a smoothing operation on the updated memory map to remove misdetection points, and perform a tiling operation on the smoothed result to store it in the data in the form of blocks.

[0070] In other embodiments of the present invention, a computer-readable storage medium containing a computer program is further provided, and when the computer program is executed by one or more processors, the memory self-updating curb storage and control method described above is executed.

[0071] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A memory self-updating curb storage and control method, characterized in that: The steps include: S1 feature preprocessing and confidence design: restore the multi-path curb detection information into 3D information, then construct and smooth the sub-images, and then design the confidence based on the confidence of the detection results themselves and the sub-image common view results; S2: Concatenate and update subgraph and memory map: If the memory map does not exist, generate the memory map using the result of the subgraph in step S1; If a memory map exists, the sub-map and the memory map are spliced ​​by nearest neighbor matching and minimum reprojection error optimization, and a search tree is constructed for the spliced ​​memory map to find the common viewing area of ​​the detection result and the memory map, triggering the confidence judgment logic and updating the memory map; S3 memory map post-processing and tiling: Smooth the updated memory map to remove false positives, and then tile the smoothed result and store it in the data in the form of blocks.

2. The method according to claim 1, characterized in that: Step S1 specifically includes: S11: Obtain visual detection information and laser 3D detection information, and perform deep recovery on the visual detection information to obtain visual 3D detection information; S12: Perform time synchronization through a separate time-triggered thread. When different threads receive visual 3D detection information and laser 3D detection information, use SLAM posture for time synchronization. The formula is as follows: Among them, X target is the coordinate value of the vehicle coordinate system at the target time, X source is the vehicle coordinate system coordinate value at source time, T target is the 4*4 transformation matrix of the SLAM pose at target time, T source It is the 4*4 transformation matrix of the SLAM pose at source time; S13: performing Ransac fitting on the synchronized visual 3D detection information and laser 3D detection information; S14: Using the SLAM pose and the result of step S13, construct a coarse sub-map with a range of N meters and a maximum number of point clouds; S15: performing MeanShift and downsampling operations on the coarse sub-image to obtain a smooth sub-image; S16: Classify the smoothed sub-image results and perform nearest neighbor search on each other, and design the confidence according to their own confidence and common viewing area.

3. The method according to claim 2, characterized in that In step S16, when performing confidence design, For a common view area, that is, when a certain area is observed by multiple detectors or perspectives, when the confidence of the detection result of at least one category is high, the overall confidence of the common view area is also high; For non-co-viewing areas, that is, areas observed by only one detector or viewpoint, the confidence provided by each detector is directly used as the confidence evaluation of the area.

4. The method according to claim 3, characterized in that In step S11, when performing depth recovery, a laser-assisted depth recovery method or an inverse projection perspective transformation IMP method is adopted.

5. The method according to claim 2, characterized in that: Step S2 specifically includes: S21: Use SLAM pose to project the sub-image in step S15 to generate a sub-image in the global coordinate system, i.e., a global sub-image, and determine whether there is a memory map at this time. If not, save the global sub-image at this time as a memory map, otherwise execute the following steps S22-S29: S22: Read the global subgraph features and construct a binary search tree of the Kd-tree point cloud for storage; S23: Classify the memory map feature points and the global feature points and perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and perform certain threshold screening, Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs; S24: Set the maximum number of iterations and the iteration termination condition. The initial pose of each iteration is the pose optimized in the previous iteration. S25: Use all matching point pairs and poses to construct the following residual: Among them, R and t are optimization variables, P lidar_i and P map_i are the global subgraph feature points and memory map feature points of the i-th matching point pair, and n represents the number of matching point pairs; S26: Use Levenberg–Marquardt to optimize the pose and minimize the residual e in step S25; S27: Repeat steps S23-S25 until the iteration condition is met, and output the optimized pose; S28: transform the coordinates of the memory map features through the optimized pose and splice them with the global sub-map; S29: Perform confidence logic judgment on the spliced ​​memory map. If the global sub-map and the initial memory map have a common view area, the one with a larger confidence is used as the update content; if there is no common view area, the spliced ​​result is used.

6. The method according to claim 1, characterized in that In step S1, the multi-path curb detection information includes periscopic camera detection information, fisheye camera detection information and laser 3D detection information.

7. The method according to claim 1, characterized in that In step S3, smoothing and noise reduction operations are completed by clustering and downsampling the updated memory map.

8. A memory self-updating road tooth storage and control device, characterized in that: Includes the following modules: Feature preprocessing and confidence design module, used to restore 3D information from multi-path curb detection information, then construct and smooth sub-images, and then design confidence based on the confidence of the detection results themselves and the sub-image common view results; The sub-graph and memory map splicing and updating module is used to generate a memory map using the results of the sub-graph in the feature preprocessing and confidence design modules when there is no memory map; when there is a memory map, the sub-graph and the memory map are spliced ​​by nearest neighbor matching and minimum reprojection error optimization, a search tree is constructed for the spliced ​​memory map, the common viewing area of ​​the detection result and the memory map is found, the confidence judgment logic is triggered, and the memory map is updated; The memory map post-processing and tiling module is used to smooth the updated memory map to remove false positives, and to tile the smoothed result and store it in the data in the form of blocks.

9. A computer-readable storage medium containing a computer program, characterized in that: When the computer program is executed by one or more processors, the memory self-updating road tooth storage and control method according to any one of claims 1 to 7 is executed.

Images