Method for efficiently generating Euclidean symbol distance field (ESDF) map based on laser radar point cloud clustering
By performing voxel downsampling, ground segmentation, and Euclidean clustering on LiDAR point clouds, combined with sparse hash tables and incremental search, the latency problem of traditional ESDF generation methods on embedded platforms is solved, achieving efficient obstacle detection and path planning.
Patent Information
- Application Number
- CN202511144080.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-15
- Publication Date
- 2025-12-02
AI Technical Summary
Traditional ESDF generation methods struggle to meet the demands of high-frequency, low-latency real-time updates on embedded or edge platforms. Furthermore, dynamic objects and noise points can cause frequent distance field diffusion, impacting overall performance and stability.
By performing voxel downsampling, ground segmentation, Euclidean clustering, and sparse hash table construction on the lidar point cloud, selective distance field calculation is performed only for obstacle regions, and incremental breadth-first search is used for local updates.
It significantly reduces computing resource consumption, supports high-frequency, low-latency real-time ESDF map building, and improves map building efficiency and stability.
Smart Images
Figure CN121049918A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot navigation and obstacle avoidance technology, specifically to a method for efficiently generating Euclidean Symbolic Distance Field (ESDF) maps based on lidar point cloud clustering. Background Technology
[0002] With the widespread application of mobile robots, autonomous driving, and drones in indoor and outdoor environments, real-time construction of high-precision 3D environment models and obstacle avoidance has become a key technology. Euclidean Signed Distance Field (ESDF), as a map representation that simultaneously provides free space and obstacle distance information, can offer rich distance constraints for path planning, collision detection, and motion control. However, traditional ESDF generation methods require distance calculations across the entire discretized mesh or voxel space. The computational load and memory consumption increase linearly or even quadratically with the environment scale, making it difficult to meet the high-frequency, low-latency update requirements of embedded or edge platforms. To improve ESDF generation efficiency, incremental update and sparse storage strategies have been proposed, but these methods still treat all hit voxels from the original point cloud (including a large amount of background noise, ground points, and dynamic targets) as "occupied" seeds, resulting in excessively large local update regions. Furthermore, dynamic objects and noise points continuously trigger frequent distance field diffusion, affecting overall performance and stability. Therefore, how to perform selective distance field calculations only on the actual obstacle areas that need to be avoided in the scene while ensuring the accuracy of ESDF distance, and further reduce the number of seeds and the spread range of incremental updates, remains a problem that current technology urgently needs to solve. Summary of the Invention
[0003] The purpose of this invention is to provide a method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, thereby reducing computational resource consumption and improving map building efficiency.
[0004] To achieve the above objectives, this invention provides a method for generating Euclidean symbolic range field maps based on lidar point cloud clustering, comprising the following steps: Step 1: Perform voxel downsampling and ground segmentation on the raw point cloud data from the lidar to obtain the preprocessed point cloud; Step 2: For the preprocessed point cloud, use the Euclidean clustering algorithm to extract and merge obstacle clusters and remove dynamic targets; Step 3: Construct a sparse hash table structure in 3D space according to the set voxel size, map the obstacle point cloud obtained in Step 2 to the corresponding voxel and mark it as occupied, and initialize the remaining voxels to an idle or unknown state. Step 4: For newly marked voxels as occupied or in a changed state, initiate an incremental breadth-first search within their neighborhood starting from the voxel, update the shortest distance value of each voxel in turn, and set the distance value of occupied voxels to negative and the distance value of idle voxels to positive.
[0005] Optionally, in the method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, step 1 of the method employs a density-based iterative voxel filtering algorithm for downsampling the original lidar point cloud. This algorithm constructs an initial voxel grid, counts the number of points within each voxel, and removes voxels with a number of points below a preset threshold and the points they contain in each iteration until no new low-density voxels are removed, thereby eliminating isolated noise and preserving high-density structures.
[0006] Optionally, in the method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, step 1 of the lidar point cloud ground segmentation process uses the Random Sample Consensus (RANSAC) algorithm. After downsampling, several points are randomly selected in the point cloud to fit a plane model, and points with a distance greater than the plane are removed according to the inlier threshold to achieve robust ground segmentation.
[0007] Optionally, in a method for generating Euclidean symbolic range field maps based on lidar point cloud clustering, step 2, obstacle extraction and dynamic target removal from the preprocessed point cloud, specifically includes: Step 1: Construct a spatial index on the preprocessed point cloud based on KD-Tree; Step 2: Call the EuclideanClusterExtraction function in the PCL library to extract clusters with preset clustering distance tolerance ε, minimum number of points N_min, and maximum number of points N_max; Step 3: Align the bounding box size with the calculation axis for each cluster, and remove clusters whose bounding box size or number of cluster points exceeds the preset range; Step 4: Calculate the centroid displacement velocity of the remaining clusters in two consecutive frames. If the velocity exceeds the threshold V_thresh, mark the corresponding cluster as a dynamic target and exclude it. Step 5: Use the final selected clusters as the obstacle set from Step 2.
[0008] Optionally, in the method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, the sparse hash table in step 3 is configured with sparse storage mode during compilation, and the initial number of buckets is pre-allocated through parameters to avoid hash collisions at runtime and improve access efficiency.
[0009] Optionally, in the method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, step 4 of the incremental breadth-first search is restricted to updating the distance within a voxel subdomain centered on the robot's current pose and extending along the X, Y, and Z directions within a preset sliding window radius. This allows for dynamic maintenance of the ESDF only in this local area, without recalculating the distance field in other areas.
[0010] This invention provides a method for generating Euclidean symbolic distance field (ESDF) maps based on lidar point cloud clustering. First, the original point cloud is downsampled and the ground is segmented and removed. Then, an Euclidean clustering algorithm is used to extract and merge obstacle point cloud clusters, while removing dynamic objects. Next, a sparse voxel hash table is initialized according to a set voxel resolution, and obstacle point clouds are mapped and marked as occupied voxels. Finally, an incremental breadth-first search algorithm is used to locally update the distance values of newly occupied or state-changed voxels. This method significantly reduces the number of voxels involved in Euclidean symbolic distance field calculations for dynamic and complex scenes, lowers memory and computational resource consumption, supports local incremental updates and sliding window mechanisms, and can achieve high-frequency, low-latency real-time ESDF map construction on embedded or edge platforms, demonstrating good engineering practical value and promising prospects for widespread application. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a flowchart illustrating a method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering according to the present invention.
[0013] Figure 2 This is a schematic diagram of the European clustering process of the present invention.
[0014] Figure 3 This is a schematic diagram of an ESDF map in an embodiment of the present invention. Detailed Implementation
[0015] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0016] Please see Figure 1 This invention provides a method for generating Euclidean symbolic distance field maps based on lidar point cloud clustering, comprising the following steps: S1: Perform voxel downsampling and ground segmentation on the raw point cloud data from the lidar to obtain the preprocessed point cloud; S2: For the preprocessed point cloud, use the Euclidean clustering algorithm to extract and merge obstacle clusters and remove dynamic targets; S3: Construct a sparse hash table structure in three-dimensional space according to the set voxel size, map the obstacle point cloud obtained in step 2 to the corresponding voxel and mark it as occupied, and initialize the remaining voxels to an idle or unknown state. S4: For a newly marked voxel as occupied or in a changed state, initiate an incremental breadth-first search within its neighborhood starting from the voxel, update the shortest distance value of each voxel in turn, and set the distance value of occupied voxels to negative and the distance value of idle voxels to positive.
[0017] The following provides further explanation in conjunction with the specific process steps; S1: Step 1 uses a density-based iterative voxel filtering algorithm to downsample the original point cloud of the lidar. This algorithm constructs an initial voxel grid, counts the number of points in each voxel, and removes voxels with fewer than a preset threshold points and the points they contain in each iteration until no new low-density voxels are removed, thereby eliminating isolated noise and preserving high-density structures.
[0018] Furthermore, the Random Sample Consensus (RANSAC) algorithm is used for ground segmentation processing of lidar point clouds. After downsampling, several points are randomly selected in the point cloud to fit a plane model, and points with a distance greater than the plane are removed according to the inlier threshold to achieve robust ground segmentation.
[0019] S2: Obstacle extraction and dynamic target removal from the preprocessed point cloud specifically include: Step 1: Construct a spatial index on the preprocessed point cloud based on KD-Tree; Step 2: Call the EuclideanClusterExtraction function in the PCL library to extract clusters with preset clustering distance tolerance ε, minimum number of points N_min, and maximum number of points N_max; Step 3: Align the bounding box size with the calculation axis for each cluster, and remove clusters whose bounding box size or number of cluster points exceeds the preset range; Step 4: Calculate the centroid displacement velocity of the remaining clusters in two consecutive frames. If the velocity exceeds the threshold V_thresh, mark the corresponding cluster as a dynamic target and exclude it. Step 5: Use the final selected clusters as the obstacle set from Step 2.
[0020] S3: Sparse hash tables enable sparse storage mode at compile time and pre-allocate the initial number of buckets through parameters to avoid runtime hash collisions and improve access efficiency.
[0021] S4: Incremental breadth-first search is restricted to updating distances within a voxel subdomain centered on the robot's current pose and extending along the X, Y, and Z directions within a preset sliding window radius. This allows for dynamic maintenance of the ESDF only in this local region, without recalculating the distance fields of other regions.
Claims
1. A method for efficiently generating Euclidean Symbolic Range Field (ESDF) maps based on lidar point cloud clustering, characterized in that, Includes the following steps: Step 1: Perform voxel downsampling and ground segmentation on the raw point cloud data from the lidar to obtain the preprocessed point cloud; Step 2: For the preprocessed point cloud, use the Euclidean clustering algorithm to extract and merge obstacle clusters and remove dynamic targets; Step 3: Construct a sparse hash table structure in 3D space according to the set voxel size, map the obstacle point cloud obtained in Step 2 to the corresponding voxel and mark it as occupied, and initialize the remaining voxels to an idle or unknown state. Step 4: For newly marked voxels as occupied or in a changed state, initiate an incremental breadth-first search within their neighborhood starting from the voxel, update the shortest distance value of each voxel in turn, and set the distance value of occupied voxels to negative and the distance value of idle voxels to positive.
2. The method for efficiently generating Euclidean Symbolic Distance Field (ESDF) maps based on lidar point cloud clustering as described in claim 1, characterized in that, Step 1 employs a density-based iterative voxel filtering algorithm for downsampling the original point cloud of the lidar. This algorithm constructs an initial voxel grid, counts the number of points within each voxel, and removes voxels with a number of points below a preset threshold and the points they contain in each iteration until no new low-density voxels are removed. This process eliminates isolated noise and preserves the high-density structure.
3. The method for efficiently generating Euclidean Symbolic Distance Field (ESDF) maps based on lidar point cloud clustering as described in claim 1, characterized in that, Step 1 employs the Random Sample Consensus (RANSAC) algorithm for ground segmentation processing of lidar point clouds. This algorithm randomly selects several points in the point cloud after downsampling to fit a plane model, and removes points whose distance from the plane is greater than a threshold based on an interior point threshold, thereby achieving robust ground segmentation.
4. The method for efficiently generating Euclidean Symbolic Distance Field (ESDF) maps based on lidar point cloud clustering as described in claim 1, characterized in that, Step 2, specifically the obstacle extraction and dynamic target removal of the preprocessed point cloud, includes: Step 1: Construct a spatial index on the preprocessed point cloud based on KD-Tree; Step 2: Call the EuclideanClusterExtraction function in the PCL library to extract clusters with preset clustering distance tolerance ε, minimum number of points N_min, and maximum number of points N_max; Step 3: Align the bounding box size with the calculation axis for each cluster, and remove clusters whose bounding box size or number of cluster points exceeds the preset range; Step 4: Calculate the centroid displacement velocity of the remaining clusters in two consecutive frames. If the velocity exceeds the threshold V_thresh, mark the corresponding cluster as a dynamic target and exclude it. Step 5: Use the final selected clusters as the obstacle set from Step 2.
5. The method for efficiently generating Euclidean Symbolic Range Field (ESDF) maps based on lidar point cloud clustering as described in claim 1, characterized in that, The sparse hash table in step 3 is configured with sparse storage mode at compile time and the initial number of buckets is pre-allocated by parameters to avoid hash collisions at runtime and improve access efficiency.
6. The method for efficiently generating Euclidean Symbolic Distance Field (ESDF) maps based on lidar point cloud clustering as described in claim 1, characterized in that, In step 4, the incremental breadth-first search is restricted to updating the distance within a voxel subdomain centered on the robot's current pose and extending along the X, Y, and Z directions within a preset sliding window radius. This allows for dynamic maintenance of the ESDF only in this local region, without recalculating the distance field in other regions.