An incremental polygonal mesh reconstruction method based on lidar point cloud data

Abstract

The present invention belongs to the technical field of real-time grid mapping in outdoor unknown environments, and specifically is an incremental polygonal grid reconstruction method based on laser radar point cloud data. The method obtains and preprocesses point cloud data, uses point cloud data to perform voxel construction and incremental grid fusion, matches new point clouds to grids, performs incremental grid division of fused and registered point clouds, and performs polygonal grid simplified mapping steps, and finally obtains a polygonal grid map that can be used in subsequent mapping processes. The present invention improves the efficiency of grid construction in complex dynamic environments through incremental grid generation, and combines point-to-grid matching, grid simplification and storage optimization methods to solve the balance problem between real-time performance, accuracy and computing resources in robot system mapping.

Description

Technical Field

[0001] The present invention belongs to the technical field of real-time grid mapping in outdoor unknown environments, and specifically is an incremental polygonal grid reconstruction method based on laser radar point cloud data. Background Art

[0002] With the improvement of machine intelligence, robots are increasingly used in high-risk tasks such as urban street fighting, post-disaster rescue and autonomous mapping. In these scenarios, robots need to perform tasks in unknown environments, and their ability to operate accurately depends on efficient environmental perception and mapping capabilities. Therefore, mobile robot autonomous mapping technology has become a research hotspot in related fields.

[0003] At present, single robots are relatively mature in the construction of 3D point cloud maps. They can generate compact and efficient environmental representations, improve the real-time performance of odometers, and facilitate their integration into downstream tasks. However, point cloud maps still have limitations. For example, they appear dense visually, but become sparse after zooming in. This requires additional processing when performing navigation and obstacle avoidance tasks, increases storage and computing burdens, and is not suitable for large-scale scenes. In contrast, environmental representation based on triangular meshes provides a better structured modeling method. This method can reduce storage redundancy, improve topological expression capabilities, and is widely used in tasks such as dense mapping and collision detection. Mesh maps can describe the environment smoothly and coherently, which helps robots navigate and make decisions in complex scenes. Therefore, efficient real-time mesh mapping has become a key direction in 3D reconstruction research.

[0004] Online incremental mesh reconstruction is crucial. First, it provides real-time preview, dynamic feedback of reconstructed areas and identification of areas to be sampled, making data collection more targeted. Second, compared with offline mapping, online reconstruction can directly generate mesh maps that can be used for navigation and analysis, thereby improving task execution efficiency. For autonomous robots, real-time updated mesh maps not only provide a more accurate representation of the environment, but also optimize path planning and improve dynamic obstacle avoidance capabilities.

[0005] However, in the simultaneous localization and mapping (SLAM) technology, real-time mesh reconstruction still faces many challenges. First, the topology rationality needs to be ensured to avoid topological breaks or reconstruction errors. Second, the algorithm needs to generate continuous and smooth triangular meshes to reduce holes and unreasonable sharp-angled patches to improve mesh stability. Finally, in the face of dynamic and complex outdoor environments, it is necessary to optimize computational efficiency and ensure the stability of incremental updates to meet the real-time navigation needs of robots in large-scale environments. In 3D reconstruction and SLAM systems, the fusion, updating and optimization of mesh maps are the key to improving mapping accuracy and adapting to dynamic environmental changes. Therefore, it is of positive significance to develop a method to improve the mapping accuracy and real-time performance of robot systems. Summary of the invention

[0006] In view of this, the present invention proposes an incremental polygonal mesh reconstruction method based on lidar point cloud data. This method uses a grid map as a structured expression of the point cloud, improves the efficiency of grid construction in complex dynamic environments through incremental grid generation, and combines point-to-grid matching, grid simplification and storage optimization methods to solve the problem of balancing the real-time performance, accuracy and computing resources of robot system mapping.

[0007] The technical solution adopted by the present invention is as follows:

[0008] An incremental polygonal mesh reconstruction method based on laser radar point cloud data comprises the following steps:

[0009] Step 1: Obtain and preprocess point cloud data to obtain radar point cloud data and IMU data during the movement of the intelligent body;

[0010] Step 2: Motion compensation for distorted point cloud: Motion compensation for point cloud distortion caused by motion within the radar frame is performed based on IMU data;

[0011] Step 3: Voxel construction and mesh incremental fusion. Using the key frame strategy, key frames are selected from the motion compensated point cloud data for sparse voxel division and mesh incremental fusion. The relationship between the vertices of the mesh in the existing voxels is updated through mesh incremental fusion to obtain the initial mesh map.

[0012] Step 4: Matching of point cloud to mesh: According to the relationship between the vertices of the mesh in the voxel, the pose estimation of the intelligent agent is calculated using the point-to-mesh ICP to achieve matching. In the registration process, the distance constraint from the point to the mesh and the normal vector similarity constraint are introduced to filter outliers.

[0013] Step 5: Incremental meshing of the fused registered point cloud, determine the radar pose according to the pose estimation obtained in step 3; transform each measurement point in the radar keyframe from the radar coordinate system to the global coordinate system according to the radar pose, and then place the points converted to the global coordinate system at the corresponding position of the grid map to obtain the registered point cloud; update the grid map with the registered point cloud based on the grid division strategy to obtain the global map;

[0014] Step 6: Simplify the polygonal mesh and build a map. For the global map, simplify the mesh by constructing an error matrix and optimizing the vertex pair shrinkage strategy, and update the simplified mesh to the global map to obtain a polygonal mesh map and a mapping odometer for subsequent mapping processes.

[0015] Furthermore, the process of motion compensation of the distorted point cloud in step 2 includes:

[0016] Step 2.1, calculate the pose transformation matrix of the intelligent body at different times according to the IMU data;

[0017] Step 2.2: For the lidar point cloud data at each moment, use the rotation matrix and acceleration data in the calculated pose transformation matrix to calculate the rotation and translation transformation of the point cloud respectively, and transform them to the starting coordinate system uniformly, so as to eliminate the point cloud distortion caused by the movement of the intelligent body and improve the accuracy of the point cloud data.

[0018] Furthermore, the process of step 3 of voxel construction and grid incremental fusion includes:

[0019] Step 3.1, for the selected key frame point cloud, divide the point cloud into voxels;

[0020] Step 3.2, traverse each point in the key frame point cloud, assign the point to the corresponding voxel, and complete the voxel division; in this process, use the voxel grid filter to downsample the divided radar point cloud for grid construction, and use the ikd-Tree algorithm to maintain the minimum distance between grid vertices during the downsampling process to avoid generating too small triangular grids;

[0021] Step 3.3: For each key frame segmentation result, incrementally merge it with the existing mesh, continuously update the connection relationship between the vertices of the mesh in the existing voxels, and thus construct an initial mesh map.

[0022] Furthermore, in step 3.3, when incrementally fusing each key frame with the existing grid, an incremental calculation method from point to voxel is adopted to ensure the accuracy of the fusion structure; at the same time, large-scale point cloud data is managed through information such as the storage location and normal vector of sparse voxels to optimize the point cloud data processing process.

[0023] Furthermore, in the registration process of step 4, the distance constraint from point to grid and the normal vector similarity constraint are introduced to filter outliers, specifically including:

[0024] In the point-to-mesh distance constraint, the point is calculated based on the mesh data using the triangle mesh vector and the point on the triangle mesh. To triangle mesh The distance is used as the filtering condition, and the specific calculation formula is: ,in, is the normal vector of the triangle mesh, is a point on the triangular mesh;

[0025] In view of the uncertainty of the normal vector direction, the absolute value of cosine similarity is used to evaluate the alignment between normal vectors; the normal vector similarity constraint is to compare the absolute value of normal vector cosine similarity with the set threshold Compare, filter matching points, set thresholds ,in Represents the normal vector of the triangular mesh, n w represents the normal vector of the constraint point, T Represents a transpose operation.

[0026] Furthermore, the incremental meshing process of the fused and registered point cloud in step 5 includes:

[0027] Step 5.1: Grid search, search the i The existing triangle mesh in the voxel, retrieve the triangle mesh from the mesh map according to the vertices in the obtained voxel, and then check whether all the vertices of these triangles are in the initial set of vertex sets of the mesh map; if so, add these triangles to the output set;

[0028] Step 5.2, mesh update, first reconstruct the triangular mesh of the newly added point cloud, then extract the existing triangular mesh from the existing mesh map, and compare it with the reconstructed triangular mesh to check whether the reconstructed triangular mesh already exists in the existing mesh. If it does, it is determined as a triangular mesh that needs to be deleted; if it does not, it is identified as a triangular mesh that needs to be updated to the voxel; in this way, the triangular mesh in the voxel is gradually updated to ensure that the triangular mesh of the voxel is correctly updated during the reconstruction process;

[0029] Step 5.3, mesh fusion, achieves mesh fusion by executing the add and delete operations of the reconstructed triangular mesh in step 5.2, and ensures that the status flags of the relevant areas are correctly synchronized.

[0030] Furthermore, the process of simplifying the polygonal mesh in step 6 includes:

[0031] Step 6.1: Grid mapping generates an initial triangular mesh, using the corresponding point cloud data in the global map obtained in step 5 as input to generate an initial triangular mesh;

[0032] Step 6.2, extracting mesh data, extracting vertex, edge and face information from the initial triangular mesh generated by mesh mapping;

[0033] Step 6.3, calculate the error matrix 𝑄, calculate the error matrix 𝑄 for each vertex, and accumulate the quadratic error based on the triangle face and its normal vector associated with each vertex;

[0034] Step 6.4: Define simplification targets and valid vertex pairs. Set the mesh simplification target according to the scene requirements, and filter out all valid vertex pairs based on edge connection relationships or set spatial thresholds.

[0035] Will The conditions for being considered a valid vertex pair are as follows:

[0036] Condition 1: Vertex With Vertex Directly connected, that is, sharing an edge;

[0037] Condition 2: The distance between two vertices ,in t is the threshold for adaptive adjustment;

[0038] Condition 3: The angle between the normal vectors of two vertices is less than the set threshold , to avoid incorrectly connecting discontinuous surfaces;

[0039] When both conditions 1 and 2 or conditions 1 and 3 are met at the same time, it can be determined that the vertex pair is a valid condition pair.

[0040] Step 6.5, iteratively perform vertex pair contraction. For each valid vertex pair, calculate the merged target vertex position to minimize the quadratic error, and then update the geometry and topology of the mesh, including removing degenerate patches and updating edge connectivity. After each contraction, recalculate the error cost associated with the affected area.

[0041] Step 6.7, generate a simplified mesh. When the simplification target is reached, the final simplified triangular mesh is obtained;

[0042] Step 6.8: Update the mesh map and update the simplified triangular mesh to the global map data structure to make it suitable for subsequent mapping process.

[0043] Furthermore, the error matrix 𝑄 in step 6.3 adopts the local curvature weighted method and is constructed in combination with the least squares fitting plane, and the process includes:

[0044] For each point v , select all connected points in its neighborhood , using the least squares method to fit the local plane , defining point v The error to the fitting plane is: ;in, a， b, c, d The best plane describing the vertex in the local area is determined by the fitting process; v x Represents the component of the vertex on the x-axis, v y Represents the component of the vertex on the y-axis, v z Represents the component of the vertex on the z-axis;

[0045] By calculating the error matrix of all vertices in the neighborhood, we can get the vertex v The global error weight is used to construct the initial error matrix, and then the curvature adjustment factor is used To dynamically adjust the error calculation: ;in, n i is the normal vector of the adjacent point, n v is the normal vector of the current vertex;

[0046] The final calculation formula of the error matrix is: ;in, p Represents a plane, is with p The set of associated vertices, represents the basic error matrix.

[0047] Due to the adoption of the above technical solution, the present invention has the following technical solutions:

[0048] (1) The present invention voxelizes the laser radar point cloud data, constructs a hierarchical data structure, and integrates the incremental mesh reconstruction method of the point-to-grid odometer to update the mesh vertices frame by frame, thereby improving data management efficiency and calculation accuracy. In this process, the point-to-grid matching method is used to optimize the pose estimation, effectively reducing the error accumulation and significantly improving the real-time performance and accuracy of mesh construction, which is particularly suitable for open outdoor environments.

[0049] (2) In the polygon mesh simplification and mapping step, the present invention constructs an error matrix to evaluate the impact of the shrinkage operation on the mesh accuracy of each vertex, optimizes the vertex shrinkage strategy, reduces the number of redundant faces and vertices, and significantly reduces the computational burden without significantly losing the mesh accuracy, thereby optimizing the mesh storage and rendering efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of an incremental polygonal mesh reconstruction method based on laser radar point cloud data in an embodiment;

[0051] Figure 2 This is an outdoor scene rendering generated by the embodiment using an incremental polygon mesh reconstruction method on the public data set Mai City;

[0052] Figure 3 The simplification rate is 0.8. Figure 2 The effect diagram after mesh simplification;

[0053] Figure 4 The simplification rate is 0.6. Figure 2 The effect diagram after mesh simplification;

[0054] Figure 5 The simplification rate is 0.6. Figure 2 The effect after mesh simplification. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0056] like Figure 1 As shown, this embodiment provides an incremental polygonal mesh reconstruction method based on laser radar point cloud data, comprising the following steps:

[0057] Step 1: Obtain and pre-process point cloud data. This embodiment uses a 360° mechanical laser radar as the main sensor for environmental perception to obtain radar point cloud data and IMU data during the movement of the intelligent body.

[0058] Step 2: Motion compensation for distorted point clouds. The laser radar will produce motion distortion during the movement of the intelligent body, which will affect feature extraction and matching, and reduce the accuracy of positioning and mapping. To solve this problem, this embodiment uses a unified coordinate system for motion compensation, calculates the pose transformation matrix through IMU data, and uses the rotation matrix and acceleration data in the calculated pose transformation matrix to calculate the rotation and translation transformation of the point cloud for the laser radar point cloud data at each moment. Finally, all point cloud data are transformed to the starting coordinate system, thereby eliminating the point cloud distortion caused by the movement of the intelligent body and improving the accuracy of the point cloud data.

[0059] Step 3: Voxel construction and mesh incremental fusion. Using the key frame strategy, key frames are selected from the motion compensated point cloud data for sparse voxel division and mesh incremental fusion to optimize storage and computing efficiency. The relationship between the vertices of the mesh in the existing voxels is updated through mesh incremental fusion to obtain the initial mesh map. The specific steps include:

[0060] Step 3.1, for the selected key frame point cloud, divide the point cloud into voxels;

[0061] Step 3.2, traverse each point in the key frame point cloud, assign the point to the corresponding voxel, and complete the voxel division; in this process, use the voxel grid filter to downsample the divided radar point cloud for grid construction, and use the ikd-Tree algorithm to maintain the minimum distance between grid vertices during the downsampling process to avoid generating too small triangular grids;

[0062] Step 3.3: For each keyframe segmentation result, incrementally fuse it with the existing grid, and continuously update the connection relationship between the vertices of the grid in the existing voxels, so as to construct an initial grid map. In order to further optimize the point cloud processing, the point-to-voxel incremental calculation method is used to ensure the efficiency and accuracy of the fusion process. When incrementally fusing each keyframe with the existing grid, this embodiment adopts the point-to-voxel incremental calculation method to ensure the accuracy of the fusion structure; at the same time, through the storage location and normal vector information of sparse voxels, large-scale point cloud data is managed to optimize the point cloud data processing process.

[0063] Step 4: Matching point cloud to mesh. According to the relationship between the vertices of the mesh in the voxel, the pose estimation of the intelligent agent is calculated using the point-to-mesh ICP to achieve matching. In the registration process, in order to ensure the correctness of the corresponding relationship, this embodiment introduces the distance constraint from point to mesh and the normal vector similarity constraint to filter outliers:

[0064] In the point-to-mesh distance constraint, the point is calculated based on the mesh data using the triangle mesh vector and the point on the triangle mesh. To triangle mesh The distance is used as the filtering condition, and the specific calculation formula is: ;in, is the normal vector of the triangle mesh, is a point on the triangular mesh. Under this constraint, the distance from the point to the mesh is required to be The distance must be less than a preset threshold This ensures that the matching points will not have a large distance deviation from the mesh surface. is the normal vector of the triangle mesh, is a point on the triangular mesh, point With triangle mesh The distance meets the condition ;

[0065] In view of the uncertainty of the normal vector direction, the absolute value of cosine similarity is used to evaluate the alignment between normal vectors; the normal vector similarity constraint is to compare the absolute value of normal vector cosine similarity with the set threshold Compare, filter matching points, set thresholds .

[0066] In order to ensure computational efficiency while improving the accuracy and robustness of pose estimation, this embodiment optimizes the pose by taking full account of the mesh topology information on the basis of obtaining the pose estimate of the intelligent body. That is, the pose of the intelligent body is optimized using an odometer method based on a point-to-mesh. During the pose optimization process, the pose is first predicted and initialized using a Kalman filter, and then the pose of the current frame is adjusted using an incremental optimization strategy. The point-to-mesh error minimization method is used to replace the traditional point-to-surface ICP algorithm. This method uses the relationship between mesh vertices to achieve fast matching. Specifically: for each triangular mesh , calculate the normal vector by the cross product between its vertices , assuming a triangular mesh The three vertices of , the normal vector calculation formula is: ;

[0067] Relative posture Represents the predicted point cloud The deviation from the global triangular mesh model. Therefore, the goal is to minimize the following point-to-mesh error: In the formula, C It represents the set of adaptation correspondences between points and grids, and the subscript 2 represents the two-norm operation.

[0068] Step 5: Incremental meshing of the fused and registered point cloud. Determine the sensor pose, i.e., the radar pose, based on the pose estimation obtained in step 3. Convert each measurement point in the radar frame from the radar coordinate system to the global coordinate system according to the radar pose, and then place the points converted to the global coordinate system at the corresponding positions of the grid map to obtain the registered point cloud. Update the grid map with the registered point cloud based on the grid division strategy to obtain the global map.

[0069] The meshing operation of this implementation refers to the process of incrementally merging the newly constructed triangular mesh Ti into the existing triangular mesh in the voxel currently saved in the map structure; the process includes:

[0070] Step 5.1: Grid search, search the i The existing triangle mesh in the voxel, retrieves the triangle mesh from the mesh map based on the vertices in the obtained voxel, and then checks whether all the vertices of these triangles are in the initial set of vertices in the mesh map; if so, these triangles are added to the output set.

[0071] Step 5.2, mesh update, first reconstruct the triangular mesh of the newly added point cloud, then extract the existing triangular mesh from the existing mesh map, and compare it with the reconstructed triangular mesh to check whether the reconstructed triangular mesh already exists in the existing mesh. If so, it is determined as a triangular mesh that needs to be deleted; if not, it is identified as a triangular mesh that needs to be updated to the voxel; in this way, the triangular mesh in the voxel is gradually updated to ensure that the triangular mesh of the voxel is correctly updated during the reconstruction process.

[0072] Step 5.3, mesh fusion, achieves mesh fusion by executing the add and delete operations of the reconstructed triangular mesh in step 5.2, and ensures that the status flags of the relevant areas are correctly synchronized.

[0073] Step 6: Simplify polygonal meshes and build maps. After obtaining the global map, simplify the mesh by constructing an error matrix and optimizing the vertex pair shrinkage strategy, and update the simplified mesh to the global map to obtain a polygonal mesh map and a map odometer for subsequent map building. This operation can not only reduce the number of vertices and meshes in the mesh, but also significantly improve the rendering capability while retaining the geometric characteristics and visual quality as much as possible. The process of simplifying polygonal meshes and building maps includes:

[0074] Step 6.1: Generate an initial triangular mesh by mapping the mesh. Use the corresponding point cloud data in the global map obtained in step 5 as input to generate an initial triangular mesh.

[0075] Step 6.2: Extract mesh data and extract vertex, edge and face information from the initial triangular mesh generated by mesh mapping.

[0076] Step 6.3, calculate the error matrix 𝑄, calculate the error matrix 𝑄 for each vertex, and accumulate and calculate the quadratic error according to the triangle face and normal vector associated with each vertex. The error matrix 𝑄 of this embodiment adopts the local curvature weighted method and combines the least squares fitting plane construction, and the process includes:

[0077] For each point v , select all connected points in its neighborhood , using the least squares method to fit the local plane , defining point The error to the fitting plane is: ;in, a, b, c, d The best plane describing the point in the local area is determined by the fitting process;

[0078] By calculating the error matrix of all vertices in the neighborhood, we can get the vertex v The global error weight is used to construct the initial error matrix, and then the curvature adjustment factor is used To dynamically adjust the error calculation: ;in, n i is the normal vector of the adjacent point, n v is the normal vector of the current vertex;

[0079] The final calculation formula of the error matrix is: ;in, p Represents a plane, is with p The set of associated vertices, represents the basic error matrix.

[0080] Step 6.4: Define simplification targets and valid vertex pairs. Set the mesh simplification target according to the scene requirements, and filter out all valid vertex pairs based on edge connectivity or by setting a spatial threshold. The conditions for determining valid vertex pairs are as follows:

[0081] Condition 1: Vertex With Vertex Directly connected, that is, sharing an edge;

[0082] Condition 2: The distance between two vertices ,in t is the threshold for adaptive adjustment;

[0083] Condition 3: The angle between the normal vectors of two vertices is less than the set threshold , to avoid incorrectly connecting discontinuous surfaces;

[0084] When both conditions 1 and 2 or conditions 1 and 3 are met at the same time, it can be determined that the vertex pair is a valid condition pair.

[0085] Step 6.5, iteratively perform vertex pair contraction. For each valid vertex pair, calculate the merged target vertex position to minimize the quadratic error, and then update the geometry and topology of the mesh, including removing degenerate faces and updating edge connectivity. After each contraction, recalculate the error cost associated with the affected area.

[0086] Step 6.7: Generate a simplified mesh. When the simplification goal is achieved, the final simplified triangular mesh is obtained.

[0087] Step 6.8: Update the mesh map and update the simplified triangular mesh to the global map data structure to make it suitable for subsequent mapping process.

[0088] Figure 2 The embodiment adopts the incremental polygon mesh reconstruction method to generate an outdoor scene rendering on the public data set Mai City; Figure 2 It can be seen that the grid map constructed by the incremental polygonal mesh reconstruction method of this embodiment can clearly observe the smooth and flat road surface and the outlines of vehicles and trees, and is effective and stable in large-scale grid mapping tasks.

[0089] Figure 3 The simplification rate is 0.8. Figure 2 The effect diagram after mesh simplification; Figure 4 The simplification rate is 0.6. Figure 2 The effect diagram after mesh simplification; Figure 5 The simplification rate is 0.6. Figure 2 The effect of mesh simplification. Figure 3-Figure 5 It can be seen that the greater the simplification rate, the more mesh vertices and triangles in the mesh map, presenting a more refined three-dimensional structure, which means that while the mesh maintains high geometric details, the storage and computational overhead is also large. As the simplification rate decreases, the simplified mesh loses some detail, but the number of vertices and triangles is significantly reduced.

[0090] From the above, it can be seen that the incremental polygonal mesh reconstruction method of this embodiment constructs a hierarchical data structure by voxelizing the lidar point cloud data, and integrates the incremental mesh reconstruction method of the point-to-mesh odometer to update the mesh vertices frame by frame, thereby improving data management efficiency and calculation accuracy. In this process, the point-to-mesh matching method is used to optimize the pose estimation, which effectively reduces the error accumulation and significantly improves the real-time and accuracy of mesh construction, especially for open outdoor environments. In the polygonal mesh simplification mapping step, by constructing an error matrix, the impact of the shrinkage operation on the mesh accuracy of each vertex is evaluated, the vertex shrinkage strategy is optimized, and the number of redundant patches and vertices is reduced. Without significantly losing the mesh accuracy, the computational burden is significantly reduced, and the mesh storage and rendering efficiency are optimized.

Claims

1. An incremental polygonal mesh reconstruction method based on lidar point cloud data, characterized in that: The following steps are involved: Step 1: Obtain and preprocess point cloud data to obtain radar point cloud data and IMU data during the movement of the intelligent body; Step 2: Motion compensation for distorted point cloud: Motion compensation for point cloud distortion caused by motion within the radar frame is performed based on IMU data; Step 3: Voxel construction and mesh incremental fusion. Using the key frame strategy, key frames are selected from the motion compensated point cloud data for sparse voxel division and mesh incremental fusion. The relationship between the vertices of the mesh in the existing voxels is updated through mesh incremental fusion to obtain the initial mesh map. Step 4: Matching of point cloud to mesh: According to the relationship between the vertices of the mesh in the voxel, the pose estimation of the intelligent agent is calculated using the point-to-mesh ICP to achieve matching. In the registration process, the distance constraint from the point to the mesh and the normal vector similarity constraint are introduced to filter outliers. Step 5: Incremental meshing of the fused registration point cloud, determine the radar pose according to the pose estimation obtained in step 3; transform each measurement point in the radar keyframe from the radar coordinate system to the global coordinate system according to the radar pose, and then place the points converted to the global coordinate system at the corresponding position of the grid map to obtain the registered point cloud; Based on the grid division strategy, the registered point cloud is used to update the grid map to obtain the global map; Step 6: Simplify the polygonal mesh and build a map. For the global map, simplify the mesh by constructing an error matrix and optimizing the vertex pair shrinkage strategy, and update the simplified mesh to the global map to obtain a polygonal mesh map for subsequent mapping process.

2. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The process of motion compensation of the distorted point cloud in step 2 includes: Step 2.1, calculate the pose transformation matrix of the intelligent body at different times according to the IMU data; Step 2.2: For the lidar point cloud data at each moment, use the rotation matrix and acceleration data in the calculated pose transformation matrix to calculate the rotation and translation transformation of the point cloud respectively, and transform them to the starting coordinate system uniformly, so as to eliminate the point cloud distortion caused by the movement of the intelligent body and improve the accuracy of the point cloud data.

3. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The process of step 3 voxel construction and grid incremental fusion includes: Step 3.1, for the selected key frame point cloud, divide the point cloud into voxels; Step 3.2, traverse each point in the key frame point cloud, assign the point to the corresponding voxel, and complete the voxel division; in this process, use the voxel grid filter to downsample the divided radar point cloud for grid construction, and use the ikd-Tree algorithm to maintain the minimum distance between grid vertices during the downsampling process to avoid generating too small triangular grids; Step 3.3: For each key frame segmentation result, incrementally merge it with the existing mesh, continuously update the connection relationship between the vertices of the mesh in the existing voxels, and thus construct an initial mesh map.

4. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 3, characterized in that: In step 3.3, when incrementally fusing each key frame with the existing grid, a point-to-voxel incremental calculation method is used to ensure the accuracy of the fused structure; At the same time, large-scale point cloud data is managed through the storage location and normal vector information of sparse voxels to optimize the point cloud data processing process.

5. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: In the registration process of step 4, the distance constraint from point to grid and the normal vector similarity constraint are introduced to filter outliers, specifically including: In the point-to-mesh distance constraint, the point is calculated based on the mesh data using the normal vector of the triangle mesh and the point on the triangle mesh. To triangle mesh The distance is used as the filtering condition, and the specific calculation formula is: ,in, is the normal vector of the triangle mesh, is a point on the triangular mesh; In view of the uncertainty of the normal vector direction, the absolute value of cosine similarity is used to evaluate the alignment between normal vectors; the normal vector similarity constraint is to compare the absolute value of normal vector cosine similarity with the set threshold Compare, filter matching points, set thresholds in, is the normal vector of the triangle mesh, n w represents the normal vector of the constraint point, T Represents a transpose operation.

6. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The incremental meshing process of fusion registration point cloud in step 5 includes: Step 5.1: Grid search, search the i The existing triangle mesh in the voxel, retrieve the triangle mesh from the mesh map according to the vertices in the obtained voxel, and then check whether all the vertices of these triangles are in the initial set of vertex sets of the mesh map; if so, add these triangles to the output set; Step 5.2, mesh update, first reconstruct the triangular mesh of the newly added point cloud, then extract the existing triangular mesh from the existing mesh map, and compare it with the reconstructed triangular mesh to check whether the reconstructed triangular mesh already exists in the existing mesh. If it does, it is determined as a triangular mesh that needs to be deleted; if it does not, it is identified as a triangular mesh that needs to be updated to the voxel; in this way, the triangular mesh in the voxel is gradually updated to ensure that the triangular mesh of the voxel is correctly updated during the reconstruction process; Step 5.3, mesh fusion, achieves mesh fusion by executing the add and delete operations of the reconstructed triangular mesh in step 5.2, and ensures that the status flags of the relevant areas are correctly synchronized.

7. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The process of step 6 polygon mesh simplification and mapping includes: Step 6.1: Grid mapping generates an initial triangular mesh, using the corresponding point cloud data in the global map obtained in step 5 as input to generate an initial triangular mesh; Step 6.2, extracting mesh data, extracting vertex, edge and face information from the initial triangular mesh generated by mesh mapping; Step 6.3, calculate the error matrix 𝑄, calculate the error matrix 𝑄 for each vertex, and accumulate the quadratic error based on the triangle face and its normal vector associated with each vertex; Step 6.4: Define simplification targets and valid vertex pairs. Set the mesh simplification target according to the scene requirements, and filter out all valid vertex pairs based on edge connection relationships or set spatial thresholds. Will The conditions for being considered a valid vertex pair are: Condition 1: Vertex With Vertex Directly connected, that is, sharing an edge; Condition 2: The distance between two vertices ,in t is the threshold for adaptive adjustment; Condition 3: The angle between the normal vectors of two vertices is less than the set threshold To avoid incorrectly connecting discontinuous surfaces; When both conditions 1 and 2 or conditions 1 and 3 are met at the same time, it can be determined that the vertex pair is a valid vertex pair; Step 6.5, iteratively perform vertex pair contraction. For each valid vertex pair, calculate the merged target vertex position to minimize the quadratic error, and then update the geometry and topology of the mesh, including removing degenerate patches and updating edge connectivity. After each contraction, recalculate the error cost associated with the affected area. Step 6.7, generate a simplified mesh. When the simplification target is reached, the final simplified triangular mesh is obtained; Step 6.8: Update the mesh map and update the simplified triangular mesh to the global map data structure to make it suitable for subsequent mapping process.

8. The incremental polygonal mesh reconstruction method based on laser radar point cloud data according to claim 7, characterized in that: The error matrix 𝑄 in step 6.3 is constructed by using the local curvature weighted method combined with the least squares fitting plane, and the process includes: For each vertex v , select all connected vertices in its neighborhood , using the least squares method to fit the local plane , the error from the definition point to the fitting plane is: ； in, a, b, c, d The best plane describing the vertex in the local area is determined by the fitting process; v x Represents the component of the vertex on the x-axis, v y Represents the component of the vertex on the y-axis, v z Represents the component of the vertex on the z-axis; By calculating the error matrix of all vertices in the neighborhood, we can get the vertex v The global error weight is used to construct the initial error matrix, and then the curvature adjustment factor is used To dynamically adjust the error calculation: ； in, n i is the normal vector of the adjacent point, n v is the normal vector of the current vertex; The final calculation formula of the error matrix is: ； Where p represents a plane, is the set of vertices associated with p, represents the basic error matrix.

Images