Vehicle high-precision positioning method and system based on laser radar point cloud data

Abstract

The invention discloses a vehicle high-precision positioning method and system based on laser radar point cloud data, and the method comprises the steps: integrating the point cloud data into a vehicle body coordinate system through timestamp synchronization and coordinate system conversion to form fused point cloud data, extracting stable features, such as a road marking and a building contour, and comparing the stable features with a high-precision map; and calculating a position deviation vector in real time. When the deviation exceeds a threshold value, a dynamic correction mechanism and an iterative nearest point algorithm are adopted for pose correction and enhancement, meanwhile, quantitative indexes such as transverse dislocation and course angle deviation are calculated through geometrical relationship analysis, and if the indexes exceed the limit, a repositioning process is triggered to adjust the pose. And finally, motion interpolation compensation is performed in combination with a vehicle dynamics model, and an error propagation chain is analyzed through the point cloud density and the matching residual error to obtain a positioning confidence index. Through multi-level deviation correction and error quantization, the positioning stability and precision in a complex environment are remarkably improved, and reliable support is provided for automatic driving.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of automatic driving, and particularly relates to a vehicle high-precision positioning method and system based on laser radar point cloud data. BACKGROUND

[0002] In the field of intelligent driving and autonomous driving, vehicle high-precision positioning technology is the core support for ensuring safety and reliability, and its importance is self-evident. Precise positioning not only concerns the accurate perception of the surrounding environment of the vehicle, but also directly affects the correctness of path planning and decision control, and is a key link for realizing the transition from theory to reality of unmanned driving. However, despite the attention paid to this field, existing technologies still face many challenges and need to be broken through. At present, many positioning methods perform poorly in complex environments, especially when facing dynamic road scenes, and it is often difficult to maintain sustained accuracy and stability. The problem is not only the limitation of the technology itself, but also the insufficient ability to comprehensively capture and process real-time environmental information. Especially in urban road or high-speed driving scenarios, the environmental information around the vehicle changes rapidly, and existing solutions are not up to the task of integrating multi-source information and responding to sudden changes, leading to deviations in positioning results and even affecting the reliability of the entire system.

[0003] In-depth analysis can find that the core technical difficulties of existing technologies are concentrated on how to effectively process and fuse a large amount of environmental perception data. Laser radar, as an important perception device, can provide rich three-dimensional spatial information, but its discontinuity in time and space becomes a major obstacle. This discontinuity will directly lead to difficulties in accurately determining the real-time position of the vehicle when the vehicle moves quickly or the environmental features are sparse. For example, in high-speed driving, the data collected by the laser radar may have intervals due to scanning frequency limitations, and the position change of the vehicle between two frames of data cannot be accurately captured, resulting in jumps or shifts in the positioning trajectory.

[0004] Therefore, how to ensure the continuity and consistency of environmental perception data in dynamic environments and eliminate positioning deviations through effective data fusion has become a key problem in the field of vehicle high-precision positioning. The solution to this problem not only needs to overcome the time discontinuity in data acquisition, but also needs to achieve precise matching and real-time correction of environmental features in complex scenarios to ensure stable positioning of vehicles under various driving conditions. SUMMARY

[0005] To solve the above technical problems, the present application provides a vehicle high-precision positioning method and system based on laser radar point cloud data. Among them, a vehicle high-precision positioning method based on laser radar point cloud data comprises:

[0006] The point cloud data collected by the multi-directional laser radar is time-stamped and converted into a coordinate system to obtain fused point cloud data;

[0007] Stable features are extracted from the fused point cloud data and compared with map reference features to obtain a position deviation vector;

[0008] The dynamic correction mechanism is triggered according to the size relationship between the position deviation vector and a preset threshold to adjust the pose, thereby obtaining a corrected pose estimate;

[0009] According to the corrected pose estimate, the continuous multi-frame point cloud data is iteratively registered in a sparse feature environment to obtain an enhanced pose estimate;

[0010] According to the enhanced pose estimate, the geometric relationship between the point cloud features and the map elements is analyzed to obtain a misplacement quantization index;

[0011] According to the size relationship between the misplacement quantization index and a preset threshold, the repositioning process is triggered to update the pose, thereby obtaining an updated pose;

[0012] According to the updated pose, the motion change is estimated and the point cloud sequence is interpolated based on the vehicle dynamics model, thereby obtaining a continuous point cloud sequence;

[0013] According to the continuous point cloud sequence, the point cloud density and the matching residual are tracked and error propagation calculation is performed, thereby obtaining a positioning reliability index and a positioning result.

[0014] Preferably, the process of obtaining the fused point cloud data comprises:

[0015] The timestamp information of the original point cloud data collected by the multi-source laser radar is analyzed to obtain the collection time of each frame of point cloud;

[0016] The time offset of each laser radar is determined according to the deviation of the collection time from the vehicle master clock;

[0017] The point cloud timestamp is corrected according to the size relationship between the time offset and a preset threshold, thereby obtaining a time-corrected point cloud;

[0018] The time-corrected point cloud is transformed into the vehicle body coordinate system according to the current pose of the vehicle and the laser radar extrinsic matrix, thereby obtaining a coordinate-unified point cloud;

[0019] The spatial overlap region of different laser radars is searched in the coordinate-unified point cloud according to the nearest neighbor search method, thereby obtaining an overlap region point cloud;

[0020] The overlap region point cloud is spatially divided by a voxel grid division method to obtain a voxelized point cloud set;

[0021] According to the voxelized point cloud set, position average calculation is performed on points of different laser radars to obtain sparse fusion point cloud data.

[0022] Preferably, the process of obtaining the position deviation vector comprises:

[0023] The sparse fusion point cloud data is preprocessed to filter out noise and irrelevant points to obtain a clean point cloud data set;

[0024] According to the clean point cloud data set, a geometric segmentation method is used to extract road markings and building outlines to obtain a stable feature set;

[0025] According to the map, reference feature data is retrieved and preliminarily registered with the stable feature set to obtain an initial matching relationship;

[0026] According to the real-time comparison method, the initial matching relationship is finely adjusted to obtain an accurate matching result;

[0027] According to the accurate matching result, the position deviation is calculated and the deviation vector data is generated;

[0028] According to the deviation vector data and the map reference feature, final verification is performed to obtain the final position deviation vector.

[0029] Preferably, the process of obtaining the corrected pose estimation comprises:

[0030] According to the position deviation vector, the modulus is calculated, and according to the size relationship between the modulus and the preset threshold, the current frame is marked as a deviation overrun state;

[0031] According to the deviation overrun state, a dynamic correction mechanism is started, and a Kalman filter is used to fuse historical pose estimation and the current deviation vector to obtain a preliminary corrected pose;

[0032] According to the least squares method, the preliminary corrected pose and the most recent multiple frame pose estimations are subjected to consistency constraint optimization to obtain an optimized corrected pose;

[0033] The optimized corrected pose is overlaid on the original pose estimation to obtain the corrected pose estimation.

[0034] Preferably, the process of obtaining the enhanced pose estimation comprises:

[0035] According to continuous multiple frame point cloud data acquisition, an initial point cloud data set is obtained;

[0036] According to the initial point cloud data set, an iterative closest point algorithm is used to preliminarily register adjacent frames to obtain a preliminarily registered point cloud relationship;

[0037] Adjust the local area point pair according to the comparison result of the preliminary registered point cloud relationship and the preset threshold, and obtain an adjusted point cloud registration result;

[0038] According to the adjusted point cloud registration result, obtain the relative pose transformation matrix between multiple frames of point clouds, and judge the continuity of the pose transformation;

[0039] According to the continuity of the pose transformation, perform global registration in a cumulative transformation manner to obtain globally consistent point cloud registration data;

[0040] According to the globally consistent point cloud registration data, calculate the final pose estimation value to obtain an enhanced pose estimation result.

[0041] Preferably, the process of obtaining the misregistration quantitative index comprises:

[0042] According to the point cloud data, extract key feature points and perform geometric matching with map elements to obtain an initial pose estimation result;

[0043] According to the initial pose estimation result, calculate the geometric relationship difference between the point cloud features and the map elements to obtain a preliminary value of lateral misregistration;

[0044] According to the preliminary value of lateral misregistration, use the least squares method to optimize the matching error to obtain lateral misregistration data;

[0045] According to the lateral misregistration data, analyze the distribution of the heading angle deviation to obtain a heading angle deviation parameter;

[0046] According to the comparison result of the heading angle deviation parameter and the preset threshold, perform local adjustment on the point cloud features to obtain a corrected deviation value;

[0047] According to the corrected deviation value and the lateral misregistration data, comprehensively deduce a misregistration quantitative parameter to obtain a misregistration quantitative index.

[0048] Preferably, the process of obtaining the updated pose comprises:

[0049] According to the current pose and map matching, calculate the misregistration quantitative error to obtain a misregistration quantitative error value;

[0050] According to the comparison result of the misregistration quantitative error value and the preset threshold, generate a trigger signal;

[0051] According to the trigger signal, start the repositioning process, and load the global map data and the current sensor observation data;

[0052] According to the particle filtering algorithm, distribute the current sensor observation data to the global map data to form a particle set and corresponding weights;

[0053] Resample the particle set and the corresponding weight, retain the high weight particle, and obtain a convergent particle set;

[0054] Select the pose corresponding to the particle with the highest weight from the convergent particle set to determine the updated pose.

[0055] Preferably, the process of obtaining a continuous point cloud sequence comprises:

[0056] According to the vehicle dynamics model, the original pose data in the collection interval is obtained and combined with the historical motion trajectory for preliminary analysis, and the initial estimation result after updating the pose is obtained;

[0057] According to the initial estimation result after updating the pose, the motion change is smoothed by using Kalman filtering method to determine the motion trajectory change trend in the collection interval;

[0058] According to the motion trajectory change trend combined with the time span of the collection interval, the missing space-time points in the point cloud sequence are predicted to obtain preliminary point cloud supplement data;

[0059] According to the preliminary point cloud supplement data, the point cloud sequence is filled by interpolation compensation technology to generate smooth transition intermediate point cloud data;

[0060] According to the intermediate point cloud data, the sequence optimization method is used to correct the consistency of the whole point cloud sequence to obtain a point cloud sequence result with higher continuity;

[0061] According to the point cloud sequence result, the local non-smooth area is dynamically adjusted and optimized to determine the final continuous point cloud sequence.

[0062] Preferably, the process of obtaining a positioning confidence index and its positioning result comprises:

[0063] According to the collection and preprocessing of the continuous point cloud sequence, the initial value of the point cloud density and the matching residual is extracted to obtain a preliminary data set;

[0064] According to the preliminary data set, the time sequence of the point cloud density is compared by using continuous analysis method, the density change mode is identified, and the significant interval of density change is determined;

[0065] According to the significant interval of density change, the influence weight of residual factor on positioning is calculated combined with the distribution of matching residual to judge the main error source;

[0066] According to the main error source, the propagation link is constructed by error propagation model to obtain the structure diagram of the propagation link;

[0067] According to the structure diagram of the propagation link, the influence degree of error propagation on the positioning confidence is analyzed, the fluctuation range of the positioning confidence is derived, and the preliminary evaluation value of the confidence is determined.

[0068] According to the preliminary evaluation value of the confidence degree and the sequence analysis result, a final value of the confidence index is calculated to obtain a quantitative result of the positioning confidence degree;

[0069] According to the comparison result of the quantitative result of the positioning confidence degree and a preset threshold value, the parameters of the error propagation model are adjusted to obtain an optimized confidence index value;

[0070] According to the optimized confidence index value, a predicted vehicle positioning result is output.

[0071] The application also provides a vehicle high-precision positioning system based on laser radar point cloud data, comprising:

[0072] A data fusion module is configured to synchronize time stamps and convert coordinate systems of point cloud data collected by a multi-directional laser radar to obtain fused point cloud data;

[0073] A feature comparison module is configured to extract stable features from the fused point cloud data and compare the features with map reference features to obtain a position deviation vector;

[0074] A dynamic correction module is configured to trigger a dynamic correction mechanism to adjust a pose according to a size relationship between the position deviation vector and a preset threshold value to obtain a corrected pose estimation;

[0075] An aggregation registration module is configured to perform iterative registration on continuous multiple frames of point cloud data in a feature sparse environment according to the corrected pose estimation to obtain an enhanced pose estimation;

[0076] A misalignment quantification module is configured to analyze a geometric relationship between point cloud features and map elements according to the enhanced pose estimation to obtain a misalignment quantification index;

[0077] A repositioning module is configured to trigger a repositioning process to update a pose according to a size relationship between the misalignment quantification index and a preset threshold value to obtain an updated pose;

[0078] An interpolation compensation module is configured to estimate motion changes and perform point cloud sequence interpolation according to the updated pose in combination with a vehicle dynamics model to obtain continuous point cloud sequences;

[0079] A positioning evaluation module is configured to track point cloud density and matching residual and perform error propagation calculation according to the continuous point cloud sequences to obtain a positioning confidence degree index and a positioning result.

[0080] Compared with the prior art, the application has the following advantages and technical effects:

[0081] This invention achieves a comprehensive improvement in vehicle positioning accuracy and stability by constructing a closed-loop processing chain of "deviation detection—dynamic correction—enhanced registration—quantitative evaluation—intelligent relocation—continuous compensation—confidence propagation". Specifically, by introducing least-squares consistency constraints for optimization, the dynamically corrected pose maintains a smooth transition across multiple frames, avoiding the abrupt changes in traditional methods. In feature-sparse environments, multi-frame clustering registration and pose continuity judgment are employed, and global consistency optimization is achieved using cumulative transformation, significantly reducing the cumulative effect of registration errors. By performing least-squares optimization and local feedback adjustment on lateral misalignment and heading angle deviation, more accurate and reliable misalignment quantification indicators are obtained. Based on the misalignment quantification indicators, the relocation process is intelligently triggered, and particle filtering and resampling mechanisms are combined to quickly converge to the optimal pose, improving relocation efficiency and accuracy. The trajectory change trend is estimated using a vehicle dynamics model, and missing spatiotemporal points are filled through Kalman filtering smoothing and interpolation compensation techniques. After sequence optimization and local dynamic adjustment, a highly smooth and spatiotemporally consistent continuous point cloud sequence is generated. By tracking point cloud density and matching residuals and constructing an error propagation chain model, the error coupling relationship and accumulation path are systematically analyzed, obtaining more objective and comprehensive positioning reliability quantification results, thus effectively supporting high-precision and reliable vehicle positioning in complex dynamic scenarios. Attached Figure Description

[0082] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0083] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0084] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0085] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0086] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0087] Example 1

[0088] like Figure 1 As shown, this embodiment provides a high-precision vehicle positioning method based on lidar point cloud data, including:

[0089] Timestamp synchronization and coordinate system conversion are performed on the point cloud data collected by the multi-directional laser radar to obtain fused point cloud data;

[0090] Stable features are extracted from the fused point cloud data and compared with map reference features to obtain a position deviation vector;

[0091] According to the size relationship between the position deviation vector and the preset threshold, a dynamic correction mechanism is triggered to adjust the pose, and a corrected pose estimate is obtained;

[0092] According to the corrected pose estimate, iterative registration is performed on continuous multiple frames of point cloud data in a sparse feature environment to obtain an enhanced pose estimate;

[0093] According to the enhanced pose estimate, the geometric relationship between the point cloud features and the map elements is analyzed to obtain a misalignment quantization index;

[0094] According to the size relationship between the misalignment quantization index and the preset threshold, a repositioning process is triggered to update the pose, and an updated pose is obtained;

[0095] According to the updated pose, the motion change is estimated and the point cloud sequence is interpolated based on the vehicle dynamics model, and a continuous point cloud sequence is obtained;

[0096] According to the continuous point cloud sequence, the point cloud density and the matching residual are tracked and error propagation calculation is performed to obtain a positioning reliability index and a positioning result.

[0097] Further, the process of obtaining the fused point cloud data includes:

[0098] The timestamp information is parsed from the original point cloud data collected by the multi-source laser radar to obtain the collection time of each frame of point cloud;

[0099] The time offset of each laser radar is determined according to the deviation of the collection time and the vehicle master clock;

[0100] According to the size relationship between the time offset and the preset threshold, the point cloud timestamp is corrected to obtain a time-corrected point cloud;

[0101] The time-corrected point cloud is transformed to the vehicle body coordinate system according to the current pose of the vehicle and the laser radar extrinsic matrix to obtain a coordinate-unified point cloud;

[0102] The spatial overlap region of different laser radars is found in the coordinate-unified point cloud according to the nearest neighbor search method to obtain an overlap region point cloud;

[0103] The overlap region point cloud is spatially divided by a voxel grid division method to obtain a voxelized point cloud set;

[0104] The position average of points from different lidars is calculated based on the voxelized point cloud set to obtain sparse fused point cloud data.

[0105] Furthermore, for example, in this embodiment, timestamp parsing is the primary step when processing multi-source LiDAR point cloud data. Assume there are three LiDAR devices, denoted as L1, L2, and L3, each acquiring point cloud data with a timestamp. Through standardized protocol parsing, the point cloud acquisition time per frame is found to be 10.2 seconds for L1, 10.3 seconds for L2, and 10.5 seconds for L3, while the vehicle's main clock displays 10.0 seconds. After comparing the discrepancies, L1 is offset by 0.2 seconds, L2 by 0.3 seconds, and L3 by 0.5 seconds. If the preset threshold is 0.4 seconds, then L3 needs correction by uniformly adding a 0.5-second offset to its point cloud timestamp, resulting in a time-corrected point cloud. This ensures that the multi-source data is aligned in the time dimension, laying the foundation for subsequent fusion.

[0106] In one possible implementation, the coordinate transformation stage relies on the vehicle pose and the extrinsic parameter matrix of the LiDAR. Assuming L1 is mounted on the top of the vehicle, L2 on the left, and L3 on the right, the extrinsic parameter matrix describes their position and orientation relative to the vehicle's body coordinate system. Through matrix transformation, the coordinates of a point in the L1 point cloud are converted from the sensor coordinate system to a new value in the vehicle coordinate system. The same transformation is performed on the L2 and L3 point clouds, ultimately forming a point cloud with unified coordinates. This step ensures spatial data consistency and avoids fusion errors caused by different coordinate systems.

[0107] For example, this embodiment uses a nearest neighbor search method to determine the overlapping area point cloud. Assuming that the L1 and L2 point clouds spatially overlap 5 meters in front of the vehicle, the closest pair of points among the point groups collected by the two devices is found within a search radius of 0.1 meters, and this is confirmed as the overlapping area point cloud. This helps identify redundant parts of multi-source data, providing a basis for subsequent optimization.

[0108] In one possible implementation, this embodiment divides the coordinate-uniform point cloud into cubic grids with sides of 0.2 meters using voxel meshing. Overlapping point clouds are assigned to the same voxels, forming a voxelized point cloud set. This processing effectively reduces the amount of data, improves computational efficiency, and preserves spatial features.

[0109] For example, in this embodiment, when calculating the positional average within each voxel, assuming a voxel contains 3 points from L1 and 2 points from L2, a single representative point is obtained by calculating the spatial positional average of these points. After all voxels are processed, sparsely fused point cloud data is formed. This method reduces redundant points, improves data consistency, and facilitates the efficient execution of subsequent perception tasks such as object detection and path planning.

[0110] In a possible implementation, the embodiment reduces the amount of fused point cloud data from the original million points to one hundred thousand points by sparsifying the fused point cloud, greatly reduces the computing resource occupation, and improves the perception accuracy by about 10% due to the alignment of time and space. This provides more reliable environmental information for the autonomous driving system and enhances safety.

[0111] Further, the process of obtaining the position deviation vector comprises:

[0112] The sparsified fused point cloud data is preprocessed to filter out noise and irrelevant points to obtain a clean point cloud dataset;

[0113] According to the clean point cloud dataset, a geometric segmentation method is used to extract road markings and building outlines to obtain a stable feature set;

[0114] According to the map, reference feature data is retrieved and preliminarily registered with the stable feature set to obtain an initial matching relationship;

[0115] According to the real-time comparison method, the initial matching relationship is finely adjusted to obtain an accurate matching result;

[0116] According to the accurate matching result, the position deviation is calculated and the deviation vector data is generated;

[0117] According to the deviation vector data and the map reference features, the final position deviation vector is obtained.

[0118] Further, for example, in the autonomous driving positioning system, after obtaining the point cloud data, the original data needs to be preprocessed to remove outliers caused by ground reflection, rain and fog interference.

[0119] Specifically, a statistical filtering method can be used, which sets that each point needs at least 30 neighborhood points within a radius of 5 meters, otherwise it is considered as noise and is removed, thereby obtaining a clean point cloud dataset. This step significantly reduces the false positive rate of subsequent matching.

[0120] In a possible implementation, when the embodiment extracts stable features from the clean point cloud using a geometric segmentation method, it prioritizes identifying road markings and building facades.

[0121] For example, ground points with a point cloud normal vector and a vertical direction included angle less than 15 degrees are clustered to extract the ground, and white markings are segmented by a reflectivity area with an intensity threshold greater than 200. At the same time, the vertical faces with a height mutation of more than 3 meters are Euclidean clustered to extract the building outlines, forming a stable feature set containing only road marking points and building edge points. This selection is because these features are long-term unchanged in high-precision maps and are least affected by the train shadow.

[0122] It should be noted that in the embodiment, when the reference features are called from the high-precision map, the pre-stored vector markings and building contour point clouds within a radius of 50 meters centered on the rough GNSS position of the vehicle are loaded. The extracted stable feature set is preliminarily registered, an initial tolerance error of 0.8 meters is set by using an intensity-weighted ICP algorithm, so as to quickly establish the corresponding relationship of the feature point pairs and lay a foundation for subsequent refinement.

[0123] For example, in the real-time comparison and refinement adjustment stage, if the residual mean after preliminary registration still reaches 0.45 meters, which exceeds the preset threshold of 0.3 meters, the re-registration mechanism is triggered. At this time, the system increases the feature weight and preferentially uses stable marking corner points in more than three consecutive frames for secondary iteration, so that the final matching error converges to 0.12 meters, and the positioning robustness is greatly improved.

[0124] In one possible implementation, in the embodiment, when the position deviation is calculated according to the accurate matching result, the map reference features are regarded as true values, and the real-time features are rotated and translated, and then the difference between the true values and the real-time features is calculated to generate a deviation vector with three components of east, north and sky. For example, the east deviation is -0.28 meters and the north deviation is 0.15 meters, and the total deviation is 0.32 meters and the azimuth angle is about 151 degrees through vector composition.

[0125] It should be noted that if the direction of the deviation vector is opposite to the cumulative drift trend within 10 seconds in the past by more than 30 degrees in the final checking stage, it is determined that it may be disturbed by temporary shielding, at which time the threshold is automatically relaxed and the iterative adjustment is started, and the rotation matrix step is gradually reduced by 0.005 degrees until the direction of the deviation vector is consistent with the expected direction, and finally a stable and reliable position deviation vector is output for vehicle pose correction. This complete process makes the positioning accuracy of the vehicle in the complex urban environment within 0.15 meters, effectively supporting the safe operation of high-level autonomous driving.

[0126] Further, the process of obtaining the corrected pose estimation includes:

[0127] Calculating the modulus of the position deviation vector, and marking the current frame as a deviation overrun state according to the size relationship between the modulus and the preset threshold;

[0128] Starting a dynamic correction mechanism according to the deviation overrun state, fusing the historical pose estimation and the current deviation vector through a Kalman filter to obtain a preliminary corrected pose;

[0129] Conducting consistency constraint optimization on the preliminary corrected pose and the latest multiple frame pose estimations according to the least square method to obtain an optimized corrected pose;

[0130] Covering the original pose estimation with the optimized corrected pose to obtain the corrected pose estimation.

[0131] Further, for example, in the process of obtaining the real-time position deviation vector, the current frame data collected by the sensor is compared with the reference data of the high-precision map to determine the position deviation of the vehicle on the road. Assuming that in the urban road environment, the vehicle obtains point cloud data through the laser radar, the system calculates that the length of the position deviation vector is 2.5 meters, and the preset threshold is 1.0 meter, which is obviously out of range. At this time, the system will mark the current frame as a deviation out-of-limit state. This marking mechanism helps to discover positioning abnormalities in a timely manner and provides a basis for subsequent correction.

[0132] For example, when the dynamic correction mechanism is started for the deviation out-of-limit state, the system dynamically adjusts the positioning parameters according to the current environment and historical data. Assuming that the vehicle is in an area with high-rise buildings, signal shielding causes positioning drift, and the dynamic correction mechanism will preferentially call historical trajectory data to combine the current deviation vector for preliminary correction. This way can effectively deal with the positioning challenge in complex environments.

[0133] For example, when the Kalman filter fuses historical pose estimation and the current deviation vector, it can be regarded as a smoothing processing method, aiming to reduce sudden errors. Assuming that the vehicle's continuous 5-frame pose estimation shows that the position deviation gradually increases, and the current deviation vector indicates a large jump deviation, the Kalman filter will integrate the weights of the two to generate a more smoothed preliminary correction pose. The advantage of this method is to avoid the dramatic fluctuation of the positioning result, and to improve the stability of the system.

[0134] For example, when the least squares method is used to perform consistency constraint optimization on the preliminary correction pose and the latest multiple frame pose estimations, it can be regarded as a process of overall adjustment of multi-source data. Assuming that the system takes the latest 10 frames of pose data and finds that 2 frames are obviously abnormal due to sensor noise, the least squares method will reduce the influence of abnormal data through optimization calculation to obtain a correction pose that is more consistent with the overall trend. The benefit of this optimization is to improve the reliability of the positioning result.

[0135] For example, when the optimized correction pose is output and the original pose estimation is covered, the system will directly apply the final correction result to the navigation module. Assuming that the vehicle deviates from the actual position by 10 meters at the intersection due to the blur of the road markings, and the optimized correction pose reduces the deviation to 0.5 meters, the navigation system can more accurately guide the vehicle after covering. The advantage of this covering mechanism is to ensure that the real-time output pose data is closer to the real position.

[0136] For example, the embodiment is directed to all the above-mentioned aspects, from the core scheme to the extended scheme, and is logically progressive. The core scheme is based on the bias vector calculation and the over-limit mark to quickly identify the problem; the extended scheme is to gradually improve the positioning accuracy through filtering and optimization means. Assuming that in long-distance high-speed driving, the core scheme can quickly capture the deviation, and the extended scheme can optimize through multiple frame data to ensure the stability of long-time driving. The benefit of this layered design is to balance real-time and accuracy, and adapt to different scene needs.

[0137] Further, the process of obtaining the enhanced pose estimation includes:

[0138] According to the continuous multi-frame point cloud data acquisition, an initial point cloud data set is obtained;

[0139] According to the initial point cloud data set, an iterative closest point algorithm is used to preliminarily register adjacent frames to obtain a point cloud relationship after preliminary registration;

[0140] According to the comparison result of the point cloud relationship after preliminary registration and the preset threshold, the local area point pair is adjusted to obtain an adjusted point cloud registration result;

[0141] According to the adjusted point cloud registration result, a relative pose transformation matrix between multiple frames of point clouds is obtained, and the continuity of the pose transformation is judged;

[0142] According to the continuity of the pose transformation, a cumulative transformation method is used for global registration to obtain globally consistent point cloud registration data;

[0143] According to the globally consistent point cloud registration data, a final pose estimation value is calculated to obtain an enhanced pose estimation result.

[0144] Further, for example, the embodiment extracts a specific implementation method from the actual application scene when processing the collection and registration task of continuous multi-frame point cloud data. For the acquisition of the initial point cloud data set, the laser radar device continuously scans the environment in a fixed scene, such as in the driving process of an autonomous vehicle, the laser radar collects 10 frames of point cloud data per second, each frame contains 100,000 points, forming an initial data set. This method can ensure the continuity and integrity of the data, laying a foundation for subsequent processing.

[0145] For example, the embodiment is directed to the use of the iterative closest point algorithm for preliminary registration, which can be understood as comparing the geometric features of adjacent frames of point clouds, finding the nearest point pair and calculating the transformation relationship. Assuming that in a certain registration, the preliminary transformation matrix between two frames of point clouds shows that the translation error is 0.5 meters and the rotation error is 2 degrees, if the preset threshold is 0.3 meters of translation and 1 degree of rotation, further adjustment is required. This preliminary registration provides a starting point for subsequent optimization.

[0146] For example, after detecting that the registration error exceeds the threshold, the local area point pair adjustment of the embodiment can be specifically implemented as selecting the area with larger error, such as the point cloud overlapping area when the vehicle turns, and reducing the error to within 0.2 meters by recalculating the matching relationship of the local point pair. This local adjustment can effectively improve the registration accuracy and ensure the accuracy of the point cloud relationship.

[0147] For example, when obtaining the relative pose transformation matrix between multiple frames of point clouds, the embodiment calculates the transformation matrix between each two frames through 5 consecutive frames of point cloud data, and judges whether the translation and rotation changes are smooth. If it is found that the translation of a certain frame increases suddenly to 1 meter, while the translation of other frames is only 0.1 meter, it can be marked as abnormal and needs to be further verified. This continuity judgment is helpful to find potential problems.

[0148] For example, for the cumulative transformation of global registration, the embodiment gradually superimposes the first 10 frames of point cloud data based on the relative transformation matrix to form the point cloud distribution in a unified coordinate system. This global consistency processing can avoid cumulative error and ensure the spatial consistency of the overall point cloud data.

[0149] For example, when calculating the final pose estimation value, the embodiment combines the global registration data to determine the pose of the vehicle at a certain time through weighted average, such as a final translation of 5.2 meters and a rotation angle of 10 degrees. This enhanced pose estimation result provides a reliable basis for subsequent navigation.

[0150] For example, the embodiment analyzes the stability index of pose estimation by analyzing the pose change amplitude of 10 consecutive frames. If the translation change is within 0.05 meters and the rotation change is within 0.5 degrees, it is considered that the stability is high. This reliability judgment can provide an important reference for system decision-making and help to improve the safety and stability of overall operation.

[0151] Further, the process of obtaining the misposition quantization index includes:

[0152] Extracting key feature points from point cloud data and performing geometric matching with map elements to obtain an initial pose estimation result;

[0153] Calculating the geometric relationship difference between the point cloud features and the map elements according to the initial pose estimation result to obtain a preliminary value of lateral misposition;

[0154] Optimizing the matching error by least squares method according to the preliminary value of lateral misposition to obtain lateral misposition data;

[0155] Analyzing the distribution of heading angle deviation according to the lateral misposition data to obtain a heading angle deviation parameter;

[0156] The point cloud features are locally adjusted according to a comparison result of the heading angle deviation parameter and a preset threshold value, to obtain a corrected deviation value;

[0157] A lateral displacement quantization parameter is derived according to the corrected deviation value and the lateral displacement data, to obtain a displacement quantization index.

[0158] Further, for example, when extracting key feature points from point cloud data, the embodiment identifies significant geometric shapes such as edges or corners in the point cloud as the basis for subsequent matching. Assuming that in the autonomous driving scenario, the point cloud data collected by the laser radar contains guardrails and road signs on both sides of the road, these structures have obvious linear or planar characteristics and can be used as key feature points. This extraction method can effectively reduce the amount of data processing while retaining important geometric information, laying the foundation for subsequent alignment.

[0159] For example, the embodiment uses a geometric matching method for preliminary alignment with map elements, which can be understood as comparing point cloud feature points with road elements in a preloaded high-precision map.

[0160] In one possible implementation, assuming that the guardrail line in the point cloud data is matched with the guardrail line in the map, the preliminary calculation results in a translation deviation of the vehicle pose of 0.8 meters and a rotation deviation of 3 degrees. This preliminary alignment can quickly locate the approximate position of the vehicle, providing a starting point for subsequent optimization.

[0161] For example, when calculating the geometric relationship difference between point cloud features and map elements, the embodiment compares the distance and angle between the two to determine the preliminary value of the lateral displacement. Assuming that in a certain calculation, the position of the road sign in the point cloud is laterally offset by 0.6 meters from the corresponding position in the map, this value can be used as the preliminary displacement data. This difference analysis helps to determine the direction and magnitude of the adjustment needed.

[0162] For example, when using the least squares method to optimize the matching error, the embodiment can make fine adjustments to the lateral displacement data. Assuming that the initial lateral displacement is 0.6 meters, after multiple iterations of optimization, the displacement value is reduced to 0.2 meters. This optimization method can significantly improve the accuracy of the matching and provide more reliable data support for subsequent analysis.

[0163] For example, when analyzing the distribution of the heading angle deviation, the embodiment can derive specific deviation parameters by statistically optimizing the matching results of the point cloud and the map. Assuming that in continuous multiple frames of data, the average value of the heading angle deviation is 2 degrees and the maximum value is 4 degrees, this distribution analysis can help determine whether the deviation is within an acceptable range.

[0164] For example, if the heading angle deviation exceeds the preset threshold, such as setting the threshold to 1 degree, local adjustment of the point cloud features is required. Assuming that in a certain frame of data, the deviation reaches 3 degrees, the deviation is corrected to 0.5 degrees by re-matching the feature points of the local area, such as road signs or lane lines. This local adjustment can effectively reduce the accumulation of deviation and improve overall consistency.

[0165] For example, when the error value is corrected and the lateral displacement data is comprehensively derived to quantify the displacement, the translation and rotation data can be combined to form the final quantization result. Assuming that the final lateral displacement is 0.2 meters and the heading angle deviation is 0.5 degrees, these parameters can be used as a reference for vehicle positioning. This comprehensive quantization method can provide more accurate input for subsequent path planning.

[0166] Further, the process of obtaining the updated pose includes:

[0167] Calculate the displacement quantization error according to the current pose and map matching, and obtain the displacement quantization error value;

[0168] Generate a trigger signal according to the comparison result of the displacement quantization error value and the preset threshold;

[0169] Start the repositioning process according to the trigger signal, load the global map data and the current sensor observation data;

[0170] According to the particle filtering algorithm, the current sensor observation data is scattered to the global map data to form a particle set and corresponding weights;

[0171] Perform resampling operation according to the particle set and corresponding weights, retain high-weight particles, and obtain a convergent particle set;

[0172] Select the pose corresponding to the particle with the highest weight from the convergent particle set to determine the updated pose.

[0173] Further, for example, when processing the displacement quantization error of the pose and map matching, the calculation of the displacement quantization error is usually based on the spatial difference between the current pose and the map data. Assuming that in a certain scenario, the vehicle is driving on an urban road, and there is a certain deviation between the point cloud data collected by the sensor and the preloaded high-precision map. By comparing the geometric features of the two, the displacement quantization error value is 0.5 meters, and the preset threshold is 0.3 meters, which is obviously beyond the limit. In this case, the system will generate a trigger signal, indicating that the pose needs to be further adjusted to ensure positioning accuracy.

[0174] For example, the repositioning process is started in response to the trigger signal, and the loading of global map data and the fusion of sensor observation data are discussed.

[0175] In one possible implementation, the system loads global map data covering the current area, while obtaining local environment information collected by sensors in real time. Assuming that the vehicle is at a complex intersection with multiple similar buildings around, the sensor data may be ambiguous. At this time, by preliminarily comparing the sensor data with the global map, the system can identify the approximate position range, providing a basis for subsequent particle filtering.

[0176] For example, the present embodiment is directed to the application of the particle filtering algorithm, including particle distribution and weight calculation. The system converts sensor observation data into multiple particles scattered in the global map, and each particle represents a possible pose hypothesis. Assuming that there are 1000 particles distributed around the intersection, the corresponding weights are calculated by comparing the matching degree of each particle with the map features. Some particles may be concentrated in the road center, with higher weights, while particles deviating from the road have lower weights. This way can gradually filter out more realistic pose hypotheses.

[0177] For example, the present embodiment is directed to the resampling operation and the generation of the convergent particle set, analyzed from the perspective of particle screening. In the resampling process, the system retains particles with higher weights and eliminates low-weight particles. Assuming that after one resampling, the number of particles is reduced to 200, and these particles are concentrated within 5 meters around the road centerline, forming a convergent particle set. This screening mechanism can effectively reduce the amount of calculation while improving the reliability of pose estimation.

[0178] For example, the present embodiment is based on the final pose update to select the pose corresponding to the particle with the highest weight. Assuming that the pose corresponding to the particle with the highest weight in the convergent particle set is 0.2 meters east of the road centerline, the system takes this pose as the updated current pose. This updating method can ensure that the vehicle still maintains high positioning accuracy in complex environments, providing a reliable basis for subsequent path planning.

[0179] In summary, the above-mentioned various links from misalignment error judgment to final pose update form a complete closed-loop process. Whether it is error comparison, trigger signal generation, or particle filtering and resampling operation, each link is closely connected and works together to improve the stability of the positioning system. Especially in complex urban environments, this process can effectively deal with the subtle differences between the map and the actual environment, ensuring the continuity and reliability of the system operation.

[0180] Further, the process of obtaining a sequence of continuous point clouds comprises:

[0181] According to the vehicle dynamics model, the original pose data within the collection interval is obtained and combined with the historical motion trajectory for preliminary analysis, to obtain an initial estimation result after updating the pose;

[0182] According to the updated initial estimation result of the pose, a Kalman filtering method is used to smooth the motion change to determine the motion trajectory change trend within the collection interval;

[0183] According to the motion trajectory change trend combined with the time span of the collection interval, the missing space-time points in the point cloud sequence are predicted to obtain preliminary point cloud supplement data;

[0184] According to the preliminary point cloud supplement data, the point cloud sequence is filled by interpolation compensation technology to generate smooth transition intermediate point cloud data;

[0185] According to the intermediate point cloud data, a sequence optimization method is used to correct the consistency of the whole point cloud sequence to obtain a point cloud sequence result with higher continuity;

[0186] According to the point cloud sequence result, the local non-smooth area is dynamically adjusted and optimized to determine the final continuous point cloud sequence.

[0187] Further, in the vehicle high-precision positioning system, the original pose data is obtained from the adjacent laser frame collection interval through the vehicle dynamics model, combined with the historical motion trajectory for preliminary analysis, and the initial estimation result after updating the pose is obtained.

[0188] Specifically, if the vehicle travels at a constant speed of 30km / h within a 0.1 second collection interval, the pose change calculated by the combination of the wheel speed meter and the IMU is 2.5 meters forward and 0.8 degrees yaw, which is used as the initial estimation to avoid sudden changes caused by using a single sensor.

[0189] In one embodiment, the initial estimation result is smoothed by extended Kalman filtering to suppress high-frequency noise and make the motion trajectory change trend within the collection interval more consistent with the actual vehicle kinematics.

[0190] It should be noted that after filtering, the yaw angle velocity fluctuation can be reduced from ±5deg / s to ±0.3deg / s or less, providing a stable basis for subsequent point cloud prediction.

[0191] It can be understood that according to the smoothed motion trajectory change trend, combined with the 0.1 second time span, the missing space-time points in the point cloud sequence caused by motion distortion are predicted.

[0192] For example, the laser radar 360-degree single-frame scanning takes 0.1 seconds, and the front point cloud has moved about 0.42 meters, while the rear point cloud is still at the old position. At this time, all points are interpolated to the end time of the frame using the uniform speed model to obtain preliminary point cloud supplement data, eliminating obvious stretching deformation.

[0193] For example, further fill the sparse area by spline interpolation compensation technology for preliminary supplementary data, when the turning radius is 50 meters and the vehicle speed is 36 km / h, the outside point cloud density decreases by about 30%, and by inserting the middle points along the motion trajectory through cubic spline, the distance between adjacent points is restored from 15 cm to about 5 cm, generating smooth transition middle point cloud data.

[0194] In an embodiment, a sequence optimization method such as sliding window bundle adjustment is used to correct the consistency of the overall point cloud sequence, and the pose residual error between adjacent frames is reduced from 0.15 meters to within 0.03 meters by jointly optimizing 10 consecutive frames of point cloud, so as to obtain a point cloud sequence result with significantly improved continuity.

[0195] Specifically, if there are local non-smooth areas on the slope or bumpy road, these areas of point cloud are optimized again through dynamic adjustment technology, for example, the local point set with z-axis fluctuation exceeding 8 cm is reweighted and iterated, so that the flatness of the final continuous point cloud sequence is improved to within 3 cm.

[0196] Preferably, the final continuous point cloud sequence is combined with the vehicle dynamics model again to perform a verification update, for example, the error between the predicted displacement and the actual wheel speed integral is less than 0.5%, so as to generate a complete data sequence with uniform time and continuous space, providing high-quality input for subsequent map construction and real-time positioning, significantly reducing cumulative drift and improving system robustness.

[0197] Further, the process of obtaining the positioning confidence index and the positioning result comprises:

[0198] According to the acquisition and preprocessing of the continuous point cloud sequence, the initial values of the point cloud density and the matching residual error are extracted to obtain a preliminary data set;

[0199] According to the preliminary data set, a continuous analysis method is used to compare the point cloud density in time sequence, identify the density change pattern, and determine the significant interval of density change;

[0200] According to the significant interval of density change, the influence weight of residual error factor on positioning is calculated in combination with the distribution of matching residual error to judge the main error source;

[0201] According to the main error source, an error propagation model is used to construct a propagation link to obtain a structure diagram of the propagation link;

[0202] According to the structure diagram of the propagation link, the influence degree of error propagation on the positioning confidence is analyzed, the fluctuation range of the positioning confidence is deduced, and the preliminary evaluation value of the confidence is determined;

[0203] According to the preliminary evaluation value of the confidence, the final value of the confidence index is calculated in combination with the sequence analysis result to obtain the quantitative result of the positioning confidence;

[0204] Adjust parameters of the error propagation model according to a comparison result of the position confidence quantification result and the preset threshold value, to obtain an optimized confidence index value;

[0205] Output a predicted vehicle positioning result according to the optimized confidence index value.

[0206] For example, in the automatic driving high-precision map construction scene, the collection of point cloud sequences is often limited by high-speed motion of vehicles and sensor frequency, resulting in uneven density distribution, which directly affects the subsequent positioning and mapping accuracy. Through the collection and preprocessing process of point cloud sequences, the original laser radar data is first voxel filtered and ground segmented, and the effective non-ground point cloud is extracted, and the point cloud density of each scanning frame and the initial value of the residual after feature matching are calculated to form a preliminary data set.

[0207] Specifically, the point cloud density is defined as the number of points in a unit spherical neighborhood, for example, within a radius of 0.5 meters. In the urban road scene, the normal frame density is usually between 150 and 200 points, while the vehicle may be accelerated or turned to below 80 points. Using a continuous analysis method, the density of the adjacent 10 frames is compared in time sequence, and it is found that the interval in which the density drops from 180 points to 90 points often corresponds to the vehicle entering a tunnel or a severely occluded area, at which time the density change pattern presents a clear step feature, which is marked as a significant interval.

[0208] In one possible implementation, in combination with the simultaneously calculated matching residual distribution, for example, the NDT matching residual mean is about 0.12 meters under normal circumstances, and the residual mean rises to 0.35 meters in the interval where the density significantly decreases. At this time, the influence weight of the residual factor on positioning is calculated by weighted analysis, and it is found that the density reduction contributes about 70% of the error increment, so it is judged that the main error source is indeed highly related to the density change.

[0209] It should be noted that once it is confirmed that the main error comes from the uneven density, the error propagation model can be constructed. For example, the density change is regarded as an input disturbance, and a propagation link is established in the form of a state space, and it is found that the error is first accumulated in the yaw angle estimation, causing the lateral positioning deviation to gradually enlarge along the vehicle driving direction, and within 20 consecutive frames, it can be accumulated from 0.1 meters to 0.8 meters, forming a clear propagation link structure diagram. Further analysis of the structure diagram can deduce the fluctuation range of the positioning confidence, for example, the confidence is stable above 0.95 in the normal section, and the error propagation peak section is reduced to 0.63, to obtain a preliminary evaluation value. Then, combined with the overall matching score of the sequence, the final positioning confidence quantification result is calculated, such as 0.78 obtained by weighted average.

[0210] Preferably, when the quantization result is lower than a preset threshold of 0.85, the system automatically triggers parameter readjustment, for example, reduces the process noise coefficient in the error propagation model to 60% of the original value, or increases the density compensation weight, regenerates the propagation link, and makes the optimized confidence index increase to 0.89, meeting the needs of subsequent fusion positioning. This confidence evaluation method based on joint analysis of density and residual error effectively avoids positioning failure caused by fluctuations in point cloud quality, and improves the robustness and continuous positioning accuracy of the system in complex urban scenarios.

[0211] Embodiment Two

[0212] As shown in Figure 2 , based on the same inventive concept, the embodiment also provides a vehicle high-precision positioning system based on laser radar point cloud data, comprising:

[0213] A data fusion module is configured to synchronize time stamps and convert coordinate systems for point cloud data collected by the multi-directional laser radar, to obtain fused point cloud data.

[0214] A feature comparison module is configured to extract stable features from the fused point cloud data and compare them with map reference features to obtain a position deviation vector.

[0215] A dynamic correction module is configured to trigger a dynamic correction mechanism to adjust the pose according to the size relationship between the position deviation vector and the preset threshold, to obtain a corrected pose estimate.

[0216] An aggregation registration module is configured to iteratively register continuous multiple frames of point cloud data in a feature sparse environment according to the corrected pose estimate, to obtain an enhanced pose estimate.

[0217] A misalignment quantization module is configured to analyze the geometric relationship between point cloud features and map elements according to the enhanced pose estimate, to obtain a misalignment quantization index.

[0218] A repositioning module is configured to trigger a repositioning process to update the pose according to the size relationship between the misalignment quantization index and the preset threshold, to obtain an updated pose.

[0219] An interpolation compensation module is configured to estimate motion changes and perform point cloud sequence interpolation according to the updated pose and a vehicle dynamics model, to obtain a continuous point cloud sequence.

[0220] A positioning evaluation module is configured to track point cloud density and matching residual error and perform error propagation calculation according to the continuous point cloud sequence, to obtain a positioning confidence index and a positioning result.

[0221] The vehicle high-precision positioning system based on laser radar point cloud data provided by the embodiment has all the advantages of the vehicle high-precision positioning method based on laser radar point cloud data provided by Embodiment One.

[0222] Embodiment three

[0223] The embodiment also discloses a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method in the embodiment one.

[0224] Embodiment four

[0225] The embodiment also discloses a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executed by a processor to implement the steps of the method in the embodiment one.

[0226] Embodiment five

[0227] The embodiment also discloses a computer program product, comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the method in the embodiment one.

[0228] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A high-precision vehicle positioning method based on lidar point cloud data, characterized in that, include: The point cloud data collected by multi-directional lidar is time-stamped and coordinate system transformed to obtain fused point cloud data; Based on the fused point cloud data, stable features are extracted and compared with map reference features to obtain a position deviation vector; The pose is adjusted by triggering a dynamic correction mechanism based on the relationship between the position deviation vector and the preset threshold, and the corrected pose estimate is obtained. Based on the corrected pose estimation, iterative registration is performed on multiple consecutive frames of point cloud data in a feature-sparse environment to obtain the enhanced pose estimation. Based on the enhanced pose estimation, the geometric relationship between point cloud features and map elements is analyzed to obtain a misalignment quantification index. Based on the relationship between the misalignment quantification index and the preset threshold, a relocation process is triggered to update the pose, thereby obtaining the updated pose. Based on the updated pose and the vehicle dynamics model, motion changes are estimated and point cloud sequence interpolation is performed to obtain a continuous point cloud sequence. Based on the continuous point cloud sequence, the point cloud density is tracked and the residual is matched, and error propagation calculation is performed to obtain the location reliability index and its positioning result.

2. The method according to claim 1, characterized in that, The process of obtaining fused point cloud data includes: The acquisition time of each frame of point cloud is obtained by parsing the timestamp information of the raw point cloud data collected by the multi-source lidar. The time offset of each LiDAR is determined based on the deviation between the acquisition time and the vehicle's master clock. The point cloud timestamps are corrected based on the relationship between the time offset and the preset threshold to obtain a time-corrected point cloud. Based on the vehicle's current pose and the lidar extrinsic parameter matrix, the time-corrected point cloud is transformed to the vehicle's body coordinate system to obtain a unified coordinate point cloud. The nearest neighbor search method is used to find the spatial overlapping region of different lidars in the unified coordinate point cloud to obtain the overlapping region point cloud. The overlapping region point cloud is spatially divided using a voxel mesh partitioning method to obtain a voxelized point cloud set. The position averaging of points from different lidars is calculated based on the voxelized point cloud set to obtain sparse fused point cloud data.

3. The method according to claim 1, characterized in that, The process of obtaining the position deviation vector includes: Preprocessing the sparse fused point cloud data to filter out noise and irrelevant points yields a clean point cloud dataset. Based on the clean point cloud dataset, road markings and building outlines are extracted using a geometric segmentation method to obtain a stable feature set; Reference feature data is retrieved from the map and initially registered with the stable feature set to obtain an initial matching relationship. The initial matching relationship is finely adjusted based on a real-time comparison method to obtain an accurate matching result; Calculate the positional deviation based on the precise matching result and generate deviation vector data; The final position deviation vector is obtained by performing a final verification based on the deviation vector data and map reference features.

4. The method according to claim 1, characterized in that, The process of obtaining the corrected pose estimate includes: The magnitude is calculated based on the position deviation vector, and the current frame is marked as having an excessive deviation state based on the relationship between the magnitude and a preset threshold. The dynamic correction mechanism is activated based on the deviation exceeding the limit state. The historical pose estimation and the current deviation vector are fused by a Kalman filter to obtain the preliminary corrected pose. The optimized corrected pose is obtained by performing consistency constraint optimization on the preliminary corrected pose and the pose estimation of the most recent multi-frames using the least squares method. The optimized pose is overlaid on the original pose estimate to obtain the corrected pose estimate.

5. The method according to claim 1, characterized in that, The process of obtaining enhanced pose estimation includes: An initial point cloud dataset is obtained by collecting point cloud data from multiple consecutive frames. Based on the initial point cloud dataset, the iterative nearest-point algorithm is used to perform preliminary registration of adjacent frames to obtain the point cloud relationship after preliminary registration. Based on the comparison results between the pre-registered point cloud relationship and the preset threshold, the point pairs in the local area are adjusted to obtain the adjusted point cloud registration result; Based on the adjusted point cloud registration results, obtain the relative pose transformation matrix between multiple frames of point clouds, and determine the continuity of the pose transformation. Based on the continuity of the pose transformation, a cumulative transformation method is used for global registration to obtain globally consistent point cloud registration data. The final pose estimate is calculated based on the globally consistent point cloud registration data to obtain the enhanced pose estimate result.

6. The method according to claim 1, characterized in that, The process of obtaining the misalignment quantification index includes: Key feature points are extracted from point cloud data and geometrically matched with map elements to obtain initial pose estimation results. Based on the initial pose estimation results, the geometric relationship difference between point cloud features and map elements is calculated to obtain a preliminary value of lateral misalignment. The matching error is optimized using the least squares method based on the preliminary value of the lateral misalignment to obtain the lateral misalignment data; Based on the analysis of the lateral misalignment data, the distribution of the heading angle deviation is obtained, and the heading angle deviation parameters are obtained. The point cloud features are locally adjusted based on the comparison result between the heading angle deviation parameter and the preset threshold to obtain the corrected deviation value; Based on the corrected deviation value and the lateral misalignment data, the misalignment quantification parameters are derived to obtain the misalignment quantification index.

7. The method according to claim 1, characterized in that, The process of obtaining the updated pose includes: The misalignment quantization error is calculated based on the current pose and the map matching, and the misalignment quantization error value is obtained. A trigger signal is generated based on the comparison result between the misalignment quantization error value and the preset threshold. The relocation process is initiated based on the trigger signal, loading global map data and current sensor observation data; The current sensor observation data is distributed to the global map data according to the particle filtering algorithm, forming a particle set and corresponding weights. A resampling operation is performed based on the particle set and its corresponding weights, retaining high-weight particles to obtain a convergent particle set. Based on the convergent particle set, select the pose corresponding to the particle with the highest weight to determine the updated pose.

8. The method according to claim 1, characterized in that, The process of obtaining a continuous point cloud sequence includes: Based on the vehicle dynamics model, the original pose data is obtained from the acquisition interval and combined with the historical motion trajectory for preliminary analysis to obtain the initial estimation result after pose update. Based on the initial estimation results after pose update, the Kalman filter method is used to smooth the motion changes and determine the trend of motion trajectory changes within the acquisition interval. Based on the trend of motion trajectory changes and the time span of the acquisition interval, the missing spatiotemporal points in the point cloud sequence are predicted to obtain preliminary point cloud supplementary data. Based on the preliminary point cloud supplementary data, the point cloud sequence is filled using interpolation compensation technology to generate smoothly transitioning intermediate point cloud data. Based on the intermediate point cloud data, a sequence optimization method is used to perform consistency correction on the overall point cloud sequence, resulting in a point cloud sequence with higher continuity. Based on the point cloud sequence results, the local non-smooth areas are dynamically adjusted and optimized to determine the final continuous point cloud sequence.

9. The method according to claim 1, characterized in that, The process of obtaining the location reliability index and its location results includes: The initial values ​​of point cloud density and matching residuals are extracted based on the acquisition and preprocessing of continuous point cloud sequences to obtain a preliminary dataset. Based on the preliminary dataset, a continuous analysis method is used to perform time-series comparison of point cloud density, identify density change patterns, and determine significant intervals of density change. Based on the significant intervals of the density changes, and combined with the distribution of the matching residuals, the influence weight of the residual factors on the positioning is calculated to determine the main sources of error. Based on the main sources of error, a propagation path is constructed using an error propagation model, resulting in a structure diagram of the propagation path. Based on the structure diagram of the propagation link, analyze the degree of impact of error propagation on the positioning signal, deduce the fluctuation range of the positioning signal, and determine the preliminary assessment value of the confidence level; The final value of the confidence index is calculated based on the preliminary assessment value of the confidence level and the results of the sequence analysis, thus obtaining the quantitative result of the location confidence index. The parameters of the error propagation model are adjusted based on the comparison between the quantitative results of the location confidence and the preset threshold to obtain the optimized confidence index value. The predicted vehicle positioning result is output based on the optimized confidence index value.

10. A high-precision vehicle positioning system based on lidar point cloud data, characterized in that, include: The data fusion module is used to synchronize the timestamps and transform the coordinate system of the point cloud data collected by multi-directional lidar to obtain fused point cloud data. The feature comparison module is used to extract stable features from the fused point cloud data and compare them with map reference features to obtain a position deviation vector; The dynamic correction module is used to trigger a dynamic correction mechanism to adjust the pose based on the relationship between the position deviation vector and the preset threshold, so as to obtain the corrected pose estimate. The clustering and registration module is used to iteratively register multiple consecutive frames of point cloud data in a feature-sparse environment based on the corrected pose estimation to obtain the enhanced pose estimation. The misalignment quantization module is used to analyze the geometric relationship between point cloud features and map elements based on the enhanced pose estimation to obtain misalignment quantization index. The repositioning module is used to trigger the repositioning process to update the pose based on the relationship between the misalignment quantification index and the preset threshold, so as to obtain the updated pose. The interpolation compensation module is used to estimate motion changes based on the updated pose and the vehicle dynamics model, and to perform point cloud sequence interpolation to obtain a continuous point cloud sequence. The positioning evaluation module is used to track the point cloud density and match the residuals based on the continuous point cloud sequence and perform error propagation calculations to obtain the positioning reliability index and its positioning results.

Images