Lidar slam map construction method and system based on hierarchical fusion

Abstract

The application relates to the technical field of laser radar mapping, and discloses a laser radar SLAM (Simultaneous Localization and Mapping) map construction method and system based on layered fusion, which constructs a coarse-fine double-layer voxel shared storage architecture, combines a time-space domain weighted fusion mechanism, a confidence linkage neighborhood search and a coarse-grained dynamic maintenance strategy, and solves core problems in the existing laser radar SLAM technology, such as the breaking of historical observation information, the imbalance of registration constraints caused by the near density and the far sparsity of the laser radar, the difficulty in coordinating the search efficiency and the noise resistance robustness, and the out-of-control of long-term running memory, etc. The high fidelity of the map geometry details is ensured through fine-grained voxels, the optimization of the calculation efficiency and the memory occupation is realized through the reference aggregation and batch maintenance of coarse-grained voxels, multi-frame observation information is fused through multi-dimensional weighted factors, the sensor noise and the accumulated pose drift are significantly inhibited, and the system can realize high-precision, high-efficiency and strong-robustness map construction in various scenes such as static, dynamic and weak texture.

Description

Technical Field

[0001] This invention relates to the field of lidar mapping technology, and in particular to a lidar SLAM map construction method and system based on hierarchical fusion. Background Technology

[0002] Simultaneous localization and mapping (SLAM) is the foundation for realizing environmental perception, path planning, and motion control of autonomous mobile robots. LiDAR has become the core sensor of SLAM systems due to its advantages such as high-precision ranging capability, strong environmental adaptability, and insensitivity to changes in illumination. In LiDAR SLAM methods based on point cloud distribution registration, the quality of the map directly determines the accuracy and robustness of pose estimation. In order to balance computational efficiency and map detail, existing systems generally adopt a voxelization strategy to organize and compress the point cloud.

[0003] For example, Chinese patent CN117906622B discloses a method, apparatus, device, and storage medium for constructing a laser point cloud map. Its core process includes the following steps: calculating the curvature of each point in a single frame of LiDAR scanning data, and selecting points with significant geometric structures as laser point cloud feature data based on preset high and low curvature thresholds; using pose data provided by an inertial sensor to perform distortion correction processing on the feature points. Specifically, by calculating the pose increment between the scanning time and the frame start time, the point is backed to a corrected position under a unified timestamp to obtain corrected laser point cloud data; downsampling the corrected feature points into a sparse point set, and combining this with the current pose prior, estimating the device's precise pose matrix in the world coordinate system using filtering or optimization methods; projecting the corrected feature points onto the world coordinate system through pose transformation. If the distance between the current device position and the previous mapping position exceeds a preset threshold, the batch of points is added to the dense laser point cloud map. Simultaneously, sparse points are also used to construct a sparse map for subsequent pose tracking.

[0004] The above-mentioned and existing related technologies often have the following drawbacks:

[0005] 1. Because LiDAR SLAM needs to frequently switch the hierarchical structure dynamically during long-term operation, and the physical characteristics of LiDAR being dense near and sparse far are not effectively controlled, the historical observation information of multi-frame spatiotemporal fusion is broken. At the same time, the registration constraint intensity is unbalanced in different areas, which ultimately leads to a continuous decline in map geometric accuracy and insufficient long-term operational stability.

[0006] 2. Because the existing LiDAR SLAM neighborhood retrieval mechanism relies on static spatial partitioning and does not dynamically bind the confidence of multi-frame fused point cloud with the index, and because there are strict resource constraints in application scenarios such as embedded platforms, a large number of low-confidence point clouds need to be traversed during retrieval. This makes it impossible to avoid single-frame noise interference and difficult to control the computational overhead of index maintenance. Ultimately, this results in a situation where retrieval efficiency and noise robustness cannot meet the actual application requirements in a coordinated manner. Summary of the Invention

[0007] The technical problem to be solved by this invention is that the existing technology has the disadvantage of not binding confidence to the structure switching of LiDAR SLAM, which makes it difficult to achieve the required accuracy and efficiency. To this end, we propose a LiDAR SLAM map construction method and system based on hierarchical fusion.

[0008] To achieve the above objectives, this application adopts the following technical solution: a lidar SLAM map construction method based on hierarchical fusion, comprising the following steps:

[0009] Step 1: Initialize the two-level voxel map parameters based on the LiDAR parameters and application scenario. The parameters include the fine-grained voxel side length. coarse-grained voxel side length Weight threshold Map boundary threshold and neighborhood search radius ;

[0010] Step 2: Construct a coarse-fine two-level voxel structure based on a hash table. The coarse-grained voxels maintain a reference list of their subordinate non-empty fine-grained voxels, while the fine-grained voxels store map point coordinates and weights to achieve index linkage.

[0011] Step 3: Perform motion distortion correction and adaptive downsampling on the original point cloud to extract representative points;

[0012] Step 4: Based on the initial pose value, transform the representative point to the world coordinate system, and use the coarse-grained voxel index and reference list to query and obtain the neighborhood point set of the representative point in the map;

[0013] Simultaneously, the neighboring point set is used to register the representative points of the current frame, and the initial pose value is optimized to obtain the accurate pose.

[0014] Step 5: Based on the precise pose, transform the representative point to the world coordinate system and query its corresponding fine-grained voxel. If the voxel is already occupied, update the map point coordinates and weights within the voxel using the spatiotemporal weighted fusion formula. The formula introduces a time decay factor to suppress dynamic noise and a spatial distribution equalization factor to compensate for the difference in point cloud density between near and far LiDAR.

[0015] Step 6: Monitor map boundaries in real time, and delete expired data in batches at the coarse-grained voxel level when the threshold is exceeded;

[0016] Step 7: Perform loop closure detection periodically. If a loop is detected, backtrack and correct the weighted update results of the map points in the most recent multi-frames based on the loop pose.

[0017] Preferably, the spatiotemporal weighted fusion updates the map point coordinates and weights within the voxel, as expressed by the following formula:

[0018] ;

[0019] ;

[0020] in, For the updated coordinates, For the updated weights, As a global coordinate system for historical convergence, The world coordinates of the new observation representative point, As the weight of historical fusion points, The time decay factor, As a spatial distribution equilibrium factor, This is the ratio coefficient of the side lengths of fine-grained and coarse-grained voxels. The confidence factor for new point observations.

[0021] Preferably, in step S2, the mapping function from coordinates to voxel indices is:

[0022] ;

[0023] in, Let V be the map origin, V be the voxel side length of the corresponding level, and M be the preset step size constant. The numbers are voxel index labels, and x, y, z are the three-dimensional coordinates of the point cloud to be added in the world coordinate system.

[0024] Preferably, in step S4, the query process for the neighborhood point set includes:

[0025] First, calculate the corresponding coarse-grained voxel index based on the representative point after the pose initial value transformation.

[0026] Then iterate through the coarse-grained voxel and its adjacent 3×3×3 neighborhood coarse-grained voxels;

[0027] Next, access the fine-grained voxels based on the reference list to obtain the corresponding map points;

[0028] Finally, spatial distance filtering is performed based on the neighborhood search radius: the Euclidean distance between each candidate neighborhood point and the projected representative point is calculated, and only points with a distance less than or equal to the neighborhood search radius are retained to form the final neighborhood point set.

[0029] Preferably, in step 6, the map boundary monitoring periodically updates the coordinate range of the points using an independent thread, and calculates the offset of the map center relative to the initial origin. and according to Batch release beyond preset map boundaries Coarse-grained voxel memory.

[0030] Preferably, in step S1, the formula for calculating the edge length of the fine-grained voxel is:

[0031] ;

[0032] in, The average spacing of the lidar point cloud;

[0033] The formula for calculating the side length of a coarse-grained voxel is:

[0034] ;

[0035] ;

[0036] in, The neighborhood search radius is N, which is the voxel granularity scaling factor. Its function is to determine the size of the coarse voxels so that the physical space span of the 3×3×3 coarse voxel neighborhood centered at any point covers the preset search distance.

[0037] Preferably, in step S3, the representative point extraction also includes an outlier removal step, which involves calculating the average distance and standard deviation of the K nearest neighbors of each candidate representative point and removing points whose distance is greater than 3 times the standard deviation.

[0038] Preferably, the spatial distribution equilibrium factor The calculation formula is:

[0039] ;

[0040] in, This represents the current point cloud density within the voxel. For global average density, It is an extremely small value.

[0041] Preferably, in step 7, the object of backtracking correction is the map points of the most recent 5 frames, and the spatiotemporal weighted fusion update of step 5 is re-executed with the optimized pose.

[0042] Preferably, the LiDAR SLAM map building system based on hierarchical fusion includes a two-level voxel map storage module, a point cloud preprocessing module, a spatiotemporal weighted update module, a confidence-linked neighborhood retrieval module, a coarse-grained dynamic maintenance module, and a map maintenance module integrated into an embedded computing platform. Each module interacts in real time through a data bus.

[0043] The confidence-linked neighborhood retrieval module is used to perform fast candidate point aggregation based on coarse-grained voxels, and further perform precise spatial distance filtering based on the neighborhood search radius to extract a set of effective neighborhood points with high confidence, supporting subsequent registration optimization.

[0044] The technical effects and advantages of this invention are as follows:

[0045] This invention combines a coarse-fine dual-level voxel shared storage architecture, a multi-dimensional spatiotemporal weighted fusion mechanism, confidence-linked neighborhood retrieval, and a coarse-grained dynamic maintenance strategy to form a full-process collaborative optimization scheme, achieving significant technical results. First, registration optimization is completed by querying the neighborhood point set based on the initial pose value. Then, map updates are performed using the precise pose, avoiding map distortion introduced by pose errors from the source. Combined with the high-fidelity storage of geometric details by fine-grained voxels and the aggregation of references by coarse-grained voxels, this solves the problems of historical observation information fragmentation and registration constraint imbalance caused by near-dense and far-sparse lidar in existing technologies. Furthermore, by fusing multi-frame observations through multi-dimensional weighted factors, sensor noise and pose drift accumulation are significantly suppressed. Simultaneously, confidence-linked retrieval achieves a synergistic improvement in retrieval efficiency and noise robustness through weighted filtering and coarse-grained candidate set selection. Coarse-grained dynamic maintenance efficiently solves the problem of long-term memory runaway, and closed-loop backtracking correction further ensures global map consistency. Attached Figure Description

[0046] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0047] Figure 1 This is a flowchart illustrating the lidar SLAM map construction method based on hierarchical fusion of the present invention.

[0048] Figure 2 This is a schematic diagram of the point cloud map construction method based on hierarchical fusion according to the present invention;

[0049] Figure 3 This is a schematic diagram of the two-level voxel map structure of the present invention. Detailed Implementation

[0050] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0051] Example 1

[0052] Reference Figure 1-3 As shown, the present invention provides a technical solution: a method and system for constructing LiDAR SLAM maps based on hierarchical fusion, comprising the following steps:

[0053] Step 1: Initialize the two-level voxel map parameters based on the LiDAR parameters and application scenario. The parameters include the fine-grained voxel side length. coarse-grained voxel side length Weight threshold Map boundary threshold and neighborhood search radius ,in:

[0054] ; ;

[0055] ;

[0056] Where N is the voxel granularity scaling factor, which determines the size of the coarse voxel so that the physical space span of the 3×3×3 coarse voxel neighborhood centered at any point covers the preset search distance.

[0057] Step 2: Construct a coarse-to-fine two-level voxel structure based on a hash table, and use a coordinate-voxel index mapping function to synchronously generate fine-grained indexes for the same spatial point. With coarse-grained index The coarse-grained voxels cover N×N×N fine-grained voxels and maintain a reference aggregation list of the subordinate non-empty fine-grained voxels. The fine-grained voxels store map point coordinates and weights, realizing storage sharing and index linkage.

[0058] Step 3: Perform voxel filtering, motion distortion correction and outlier removal on the original point cloud, extract representative points, and suppress uneven distribution caused by near-dense and far-sparse points;

[0059] Step 4: Based on the initial pose value provided by the odometry, transform the representative point to the world coordinate system to obtain its projected position;

[0060] Calculate the corresponding coarse-grained voxel index using the projected position coordinates;

[0061] Traverse all coarse-grained voxels within the index and its three-dimensional neighborhood (3×3×3);

[0062] By using the reference list maintained by each coarse-grained voxel, the map points stored in all associated fine-grained voxels are aggregated to form an initial candidate neighborhood point set.

[0063] Based on the initial neighborhood search radius R, the initial candidate point set is precisely spatially filtered: the Euclidean distance between each candidate point and the projected position of the representative point is calculated, and all points with a distance greater than R are eliminated, thereby obtaining a precise and spatially compact neighborhood point set.

[0064] Using this neighborhood point set as the registration target, the representative points of the current frame are registered, the initial pose value is optimized, and the accurate pose is obtained.

[0065] Step 5: Based on the precise pose, transform the representative point to the world coordinate system and query its corresponding fine-grained voxel. If the voxel is already occupied, update the map point coordinates and weights within the voxel using the spatiotemporal weighted fusion formula. The formula introduces a time decay factor to suppress dynamic noise and a spatial distribution equalization factor to compensate for the difference in point cloud density between near and far LiDAR.

[0066] Step 6: During the map boundary maintenance phase, monitor the map extent in real time; if it exceeds... Calculate the origin offset in real time and delete expired data in batches using coarse-grained voxels as the unit;

[0067] Step 7: Perform loop closure detection periodically. If a loop is detected, backtrack and correct the weighted update results of the map points in the most recent multi-frames based on the loop pose.

[0068] In a preferred embodiment, the coordinate-voxel index mapping function is:

[0069] ,in, The map origin is V, the voxel side length of the corresponding level is V, and M is a preset step size constant to ensure that the index conflict rate is ≤0.5%. The numbers are voxel index labels, and x, y, z are the three-dimensional coordinates of the point cloud to be added in the world coordinate system.

[0070] In a preferred embodiment, the spatiotemporal weighted fusion formula is as follows:

[0071] ;

[0072] ;

[0073] in: For the updated coordinates, The updated weights;

[0074] (Time decay factor): 0.99 for static scenes and 0.95 for dynamic scenes, used to balance the weights of new and old observations;

[0075] (Spatial distribution equilibrium factor): , For the current voxel density, For global average density, The value is extremely small, compensating for the difference between near and far density;

[0076] (Hierarchical coefficient): To ensure information synchronization during level switching;

[0077] (Observation confidence factor): , Let K be the standard deviation of the nearest neighbors. Adapt to new quality standards.

[0078] In a preferred embodiment, during point cloud preprocessing, the voxel filter side length = 0.5 × The center of gravity of the voxel is retained as the representative point;

[0079] Motion distortion correction based on IMU pose increment The formula is ;

[0080] Outlier removal is performed using K=10 nearest neighbor analysis, removing outliers with a distance > point.

[0081] A lidar SLAM system based on a two-level voxel structure and spatiotemporal weighted updates is used to implement the above method, including integration into an embedded computing platform:

[0082] Two-level voxel map storage module: Constructs a hierarchical structure, stores fine-grained point cloud data and coarse-grained reference aggregation list, and performs coordinate-voxel index mapping;

[0083] Point cloud preprocessing module: performs voxel filtering, distortion correction and outlier removal, and outputs representative points;

[0084] Spatiotemporal weighted update module: performs representative point pose transformation and weighted fusion update, and maintains single-point voxel constraints;

[0085] Confidence-based neighborhood retrieval module: Enables coarse-grained candidate set filtering and high-confidence point set extraction, supporting distribution registration;

[0086] Coarse-grained dynamic maintenance module: Monitors map boundaries and performs origin offset and batch deletion of expired data;

[0087] Closed-loop optimization module: Optimizes pose and backtracks to update the map, improving long-term consistency.

[0088] Example 2

[0089] Reference Figure 1-3 As shown, this embodiment is applied to an indoor warehousing scenario, and the robot is equipped with a 16-line LiDAR (average spacing of point cloud). The scenario mainly involves static shelves with no obvious dynamic interference. The core requirements are high-precision mapping and low memory usage.

[0090] Step 1: Parameter initialization:

[0091] Fine-grained voxel edge length ;

[0092] Neighborhood search radius coarse-grained voxel side length ( );

[0093] Weight threshold Map boundary threshold The hash table load factor is 0.7.

[0094] Step 2: Construction of the two-level voxel structure:

[0095] In the index mapping function, the map origin , ;

[0096] At the edge of the shelf For example:

[0097] Fine-grained index

[0098] ;

[0099] coarse-grained index

[0100] ;

[0101] The coarse-grained voxel 5003001 maintains a reference aggregation list of 8×8×8=512 fine-grained voxels. After the fine-grained voxel corresponding to point P is occupied, its index 40030010 is automatically added to this list.

[0102] Step 3: Point cloud preprocessing:

[0103] The voxel filter side length is 0.5 × 0.08 = 0.04 m, which downsamples 30,000 original points per frame to 12,000 representative points.

[0104] Motion distortion correction: The robot moves at a speed of 0.5 m / s, and the time difference between the scanning time of a certain point and the start time of the frame is... Translation vector Correction formula ;

[0105] Outlier removal: K=10 (average nearest neighbor distance) , Noise points with a distance >0.03m are removed.

[0106] Step 4: Neighborhood search and pose optimization:

[0107] Using the transformed representative point coordinates, calculate the corresponding coarse-grained voxel index, traverse the index and the coarse-grained voxels in the 3×3×3 neighborhood, and obtain all associated candidate map points by referencing the aggregation list.

[0108] Distance filtering is performed based on the neighborhood search radius R=0.5m. The distance between each candidate point and the representative point is calculated, and only points with a distance less than or equal to 0.5m are retained to form a precise neighborhood point set for registration.

[0109] Using this precise neighborhood point set as the target, the ICP algorithm is used to register the representative point of the current frame, and the initial pose value is optimized into a precise pose.

[0110] Step 5: Spacetime Weighted Update

[0111] Representative point pose transformation: initial pose Transformation formula ;

[0112] Fusion computing: A fine-grained voxel has stored historical points. , Static scene Current voxel density Global average density , , New Point , ;

[0113] ;

[0114] .

[0115] Step 6: Coarse-grained dynamic maintenance:

[0116] After running for 2 hours, the map area Maintenance was not triggered; after running for 3 hours, Calculate the offset The origin is shifted to (0.5,0,0), and batch deletion exceeds the limit. Bold pixels in the range, 0.5MB of memory reclaimed.

[0117] Step 7: Closed-loop optimization:

[0118] The pose is optimized based on the neighborhood point set, and the positioning accuracy is corrected to within 2cm.

[0119] After the robot moves 6m, it triggers closed-loop detection, corrects the drift to 0.02m, and re-integrates the representative points from the most recent 5 frames, reducing the map deviation to 0.01m.

[0120] Example 3

[0121] Reference Figure 1-3 As shown, this embodiment is applied to an autonomous driving scenario on urban roads, and the vehicle is equipped with a 64-line LiDAR (average point cloud spacing). The scenarios include dynamic targets such as vehicles and pedestrians, and the core requirements are dynamic environment adaptability and real-time performance.

[0122] Step 1: Parameter initialization:

[0123] Fine-grained voxel edge length ;

[0124] Neighborhood search radius coarse-grained voxel side length ( );

[0125] Weight threshold Map boundary threshold The hash table load factor is 0.7.

[0126] Step 2: Construction of the two-level voxel structure:

[0127] Index mapping: Road street light points ,

[0128] ;

[0129] ;

[0130] The coarse-grained voxel 30015009 maintains a reference aggregation list of 13×13×13=2197 fine-grained voxels, and the fine-grained voxel index corresponding to point Q is added to this list.

[0131] Step 3: Point cloud preprocessing:

[0132] With a voxel filter side length of 0.02m, the number of point clouds is reduced to 20,000 per frame after downsampling.

[0133] Motion distortion correction: Vehicle speed 36km / h (10m / s), at a certain point Translation vector After correction, the ghosting is eliminated;

[0134] Outlier removal: , Points with a distance greater than 0.024m are removed.

[0135] Step 4: Neighborhood search and pose optimization:

[0136] Based on the initial odometry pose value, the representative point is transformed to the world coordinate system;

[0137] Traverse the coarse-grained voxels and 3×3×3 neighborhoods corresponding to the representative point, and obtain all associated candidate map points through the reference list;

[0138] Based on the neighborhood search radius R=1.0m, distance filtering is performed to remove candidate points that are more than 1.0m away from the representative point, thus forming the final neighborhood point set;

[0139] Registration is performed using this neighborhood point set to obtain the accurate pose.

[0140] Step 5: Spatiotemporal weighted update:

[0141] Dynamic Scene A certain historical point , Current voxel density Global average density , , New Point , ;

[0142] ;

[0143] ;

[0144] Step 6: Coarse-grained dynamic maintenance:

[0145] After driving 10km, Calculate the offset The origin is shifted to (2.5,0,0), and batch deletion exceeds the limit. Bold pixels in the range, 3.2MB of memory was reclaimed.

[0146] Step 7: Closed-loop optimization:

[0147] After pose optimization, the accuracy was reduced to 1.8cm, the loop closure detection drift was corrected to 0.03m, and the most recent 5 frames were re-fused, improving the dynamic adaptability of the map.

[0148] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A lidar SLAM map construction method based on hierarchical fusion, characterized in that, Includes the following steps: S1: Initialize the two-level voxel map parameters based on the LiDAR parameters and application scenario. The parameters include fine-grained voxel side lengths. coarse-grained voxel side length Weight threshold Map boundary threshold and neighborhood search radius ; S2: Construct a coarse-fine two-level voxel structure based on a hash table. The coarse-grained voxels maintain a reference list of their subordinate non-empty fine-grained voxels, while the fine-grained voxels store map point coordinates and weights to achieve index linkage. In step S2, the mapping function from coordinates to voxel indices is: ； in, Let V be the map origin, V be the voxel side length of the corresponding level, and M be the preset step size constant. Here are the voxel index labels, and x, y, z are the three-dimensional coordinates of the point cloud to be added in the world coordinate system. S3: Perform motion distortion correction and adaptive downsampling on the original point cloud to extract representative points; S4: Based on the initial pose value, transform the representative point to the world coordinate system. Using the coarse-grained voxel index and reference list, query and obtain the neighborhood point set of the representative point in the map. At the same time, use the neighborhood point set to register the representative point of the current frame and optimize the initial pose value to obtain the accurate pose. S5: Based on the precise pose, transform the representative point to the world coordinate system, query its corresponding fine-grained voxel, and if the voxel is already occupied, update the map point coordinates and weights within the voxel through spatiotemporal weighted fusion; the formula for updating the map point coordinates and weights within the voxel through spatiotemporal weighted fusion is as follows: ； ； in, For the updated coordinates, For the updated weights, As a global coordinate system for historical convergence, The world coordinates of the new observation representative point, As the weight of historical fusion points, The time decay factor, As a spatial distribution equilibrium factor, This is the ratio coefficient of the side lengths of fine-grained and coarse-grained voxels. Confidence factor for new point observations S6: Real-time monitoring of map boundaries; when the threshold is exceeded, expired data is deleted in batches at the coarse-grained voxel level. S7: Perform loop closure detection periodically. If a loop is detected, backtrack and correct the weighted update result of the map points in the most recent multiple frames based on the loop pose.

2. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: In step S4, the query process for the neighborhood point set includes: First, calculate the corresponding coarse-grained voxel index based on the representative point after the pose initial value transformation. Then iterate through the coarse-grained voxel and its adjacent 3×3×3 neighborhood coarse-grained voxels; Next, access the fine-grained voxels based on the reference list to obtain the corresponding map points; Finally, spatial distance filtering is performed based on the neighborhood search radius: the Euclidean distance between each candidate neighborhood point and the projected representative point is calculated, and only points with a distance less than or equal to the neighborhood search radius are retained to form the final neighborhood point set.

3. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: In step S6, map boundary monitoring periodically updates the coordinate range of points using an independent thread, and calculates the offset of the map center relative to the initial origin. and according to Batch release beyond preset map boundaries Coarse-grained voxel memory.

4. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: In step S1, the formula for calculating the edge length of fine-grained voxels is: ； in, The average spacing of the lidar point cloud; The formula for calculating the side length of a coarse-grained voxel is: ； ； in, Where is the neighborhood search radius, and N is the voxel granularity scaling factor.

5. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: In step S3, the representative point extraction also includes an outlier removal step, which involves calculating the average distance and standard deviation of the K nearest neighbors of each candidate representative point and removing points whose distance is greater than 3 times the standard deviation.

6. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: The spatial distribution equilibrium factor The calculation formula is: ； in, This represents the current point cloud density within the voxel. For global average density, It is an extremely small value.

7. The lidar SLAM map construction method based on hierarchical fusion according to claim 1, characterized in that: In step S7, the object of backtracking correction is the map points of the most recent 5 frames. The spatiotemporal weighted fusion update of step S5 is re-executed with the optimized pose.

8. A lidar SLAM map building system based on hierarchical fusion, implementing the method of any one of claims 1-7, characterized in that, It includes a two-level voxel map storage module, a point cloud preprocessing module, a spatiotemporal weighted update module, a confidence-linked neighborhood retrieval module, a coarse-grained dynamic maintenance module, and a map maintenance module, all integrated into an embedded computing platform. These modules interact in real time through a data bus.

Images