Clustering point cloud-driven underground parking lot plane structure extraction and updating method

Abstract

The invention relates to a method for extracting and updating a plane structure of a parking lot, in particular to a method for extracting and updating a plane structure of an underground parking lot based on clustering LIDAR point cloud. The working process of the method is divided into two stages: plane structure extraction and plane structure updating. In the plane structure extraction stage, for the problem of heterogeneity of the wall structure of the underground parking lot environment, accurate clustering of the wall point cloud is realized based on a two-stage DBSCAN algorithm; in addition, a column point cluster denoising algorithm based on Gaussian kernel density estimation and a straight line segment extraction algorithm based on path search are introduced in the method, so that the problem that the extraction precision of the plane structure is reduced due to wall point cloud quality degradation is solved. In the plane structure updating stage, cross-source registration of LiDAR and DXF is achieved through feature consistency constraint. In order to solve the problem of wall point cloud overall offset caused by insufficient SLAM mapping precision, an overall offset sensing algorithm based on adjacency clustering identifies and calibrates an offset structure, and misjudgment in update detection is avoided.

Description

Technical Field

[0001] The present invention relates to a method for accurately extracting and updating the plane structure in an underground parking environment, specifically a method for extracting and updating the plane structure of an underground parking lot based on clustered 3D laser radar point clouds. Background Art

[0002] With the rapid growth of urban car ownership, parking difficulties are becoming increasingly prominent. The development of smart parking systems has become a key approach to addressing urban parking issues. One of the core technologies for smart parking in cities is the construction of parking lot maps. However, traditional methods for building parking lot maps suffer from high labor costs and low update efficiency, making them inadequate for the practical needs of smart parking systems. To address these issues, researchers have begun to focus on automatic generation of parking lot maps from DXF (Drawing Exchange Format) files for parking lot navigation. However, such methods typically require preprocessing of DXF files and face numerous challenges in practical application. For example, DXF files may be missing for some older parking lots, or the layout of parking lots may change over time, requiring updates to existing floor plans. Relying solely on high-precision manual measurement and hand-drawn updates is not only costly but also inefficient. Therefore, despite the potential application value of automatic generation methods, their widespread adoption remains challenging.

[0003] This patent relies on the high-precision spatial structure perception capability of LiDAR (Light Detection and Ranging) point clouds to conduct research on the automated extraction and updating of parking lot planar structures.

[0004] In terms of planar structure extraction, Kong et al. [1] The point cloud data is projected onto a plane and its boundary is extracted. The plane is then divided into a grid, and the sliding window technique is used to extract ordered corner points. Finally, these corner points are connected to generate a floor plan. Gao et al. [2] The point cloud data of the wall is projected onto a two-dimensional plane, and then the line segments in the two-dimensional point cloud are detected and segmented using an iterative RANSAC (Random Sample Consensus) clustering algorithm. Finally, the topological structure of the discrete line segment set is restored through this process. For plane structure extraction in a high-noise environment, Liu et al. [3]A line growing method based on principal component analysis (PCA) is proposed, combined with a filtering step to improve robustness. However, these methods all rely on extracting local geometric features and restoring topological relationships, and their effectiveness is highly dependent on the quality of the input point cloud. In large, noisy underground parking environments, the thick wall point cloud makes these planar structure extraction methods based on short line segment extraction and fusion perform poorly, making it difficult to generate accurate and connected planar structures.

[0005] In terms of planar structure updating, He et al. [4] The MAiD (Machine-Assisted iD) tool was proposed to optimize the mapping process by integrating automatically inferred road segments into the OpenStreetMap iD editor and verifying them with human users. However, this method is still a semi-automated mode, and its update process is highly dependent on human users to verify the automatically inferred road segments, which not only leads to limited efficiency but also may introduce subjective bias, affecting the objectivity and consistency of the results. Bastani et al. [5] The proposed fusion algorithm extends the plane structure extraction algorithm to the plane structure update task. This fusion algorithm updates the plane structure by fusing the extraction results with the existing plane structure, but it cannot detect the plane structure that exists in the original plane image but does not exist in the ground truth image. Zhou et al. [6] The UGRoadUpd framework was proposed to guide road updates using unchanged roads. This framework improves the quality of the updated road network by limiting the scope of road updates and learning features from unchanged roads. However, the aforementioned planar structure update methods often rely on road network heuristics, which limits their applicability in underground parking scenarios due to the weak correlation between road networks.

[0006] [1]Kong Q, Liao L, Yuan C. Rapid generation of editable engineering drawings from 3D point cloudreconstruction for large-scale buildings[J]. Journal of Building Engineering, 2023,63:105486.

[0007] [2]Gao X,Yang R.Floor plan reconstruction fromindoor 3D point cloudsusing iterative RANSAC linesegmentation[J].Journal of Building Engineering,2024,89:109238.

[0008] [3]Liu X,Bondarev E.Floor-plan generation from noisy point clouds[C].Fifteenth International Conference onMachine Vision,Rome,Italy,2023,305-314..

[0009] [4]Bastani F,He S,Abbar S,et al.Machine-assisted map editing[C].Proceedings of the 26th ACM SIGSPATIALInternational Conference on Advancesin Geographic Information Systems,New York,USA,2018,23-32.

[0010] [5]Bastani F,Madden S.Beyond road extraction:A dataset for map updateusing aerial images[C].IEEE / CVFInternational Conference on Computer Vision,Montreal,Canada,2021,11905-11914.

[0011] [6] Zhou M, Sui H, Chen S, et al.UGRoadUpd: An unchanged-guided historicalroad database updatingframework based on bi-temporal remote sensing images[J]. IEEE Transactions on Intelligent TransportationSystems, 2022, 23(11):21465-21477. Summary of the Invention

[0012] During the plane structure extraction phase, the present invention uses the pre-processed LiDAR wall point cloud as input, performs preliminary clustering of the point cloud data using a two-stage DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, and then performs a refined analysis of each point cluster based on the geometric distribution characteristics of the point cloud, thereby achieving accurate distinction between wall point clusters and column point clusters. Furthermore, a column point cluster denoising algorithm based on Gaussian kernel density estimation is proposed to restore the unstructured clumping distribution of column point clusters to a structured distribution, thereby improving the accuracy of column plane structure extraction. In addition, based on the morphological skeleton extraction results, a straight line segment extraction algorithm based on path search is proposed. While ensuring extraction accuracy, it significantly improves the smoothness of the plane structure and outputs the final parking lot plane structure.

[0013] In the plane structure update stage, the present invention uses the plane structure extracted by LiDAR point cloud With the pre-acquired DXF plane structure G (as the original plane structure) as input, the cross-source registration between LiDAR and DXF is first achieved through feature consistency constraints, and alignment is performed. The scale and posture of G are then updated. Subsequently, the columns and wall structures are preliminarily updated and tested, and the overall offset is perceived based on the neighbor clustering algorithm to identify the update area that may be misjudged. Furthermore, the plane structure pose is optimized through iterative closest point (ICP) registration and a secondary detection is performed. Finally, the remaining wall and column structures are updated and the updated plane structure is output.

[0014] To achieve the purpose of this invention, the following technical solution is adopted: a clustered point cloud driven underground parking lot planar structure extraction and update method. It includes the following steps:

[0015] S1. Classification of wall column point clusters:

[0016] S1-1. Point cloud clustering based on two-stage DBSCAN. The original wall point cloud is subjected to preliminary DBSCAN classification using coarse-grained parameter selection to obtain the initial clustering results. For each initial cluster C i Further DBSCAN secondary classification is performed using fine-grained parameter selection. At the same time, the point cluster with the largest number of points in the newly generated point cluster is recorded. If the number of point clusters after secondary classification n is less than 2, that is, the fine-grained DBSCAN classification result is still a single cluster, then the initial point cluster C i As the final classification result. Otherwise, traverse all newly generated point clusters like The number of points in is less than the set threshold N fined , or its point cloud distribution shows linear characteristics (determined by PCA), then the point cluster is determined to be a small cluster generated by over-segmentation, and it is combined with the largest new point cluster Merge. On the contrary, if If the above conditions are not met, it is determined to be a valid point cluster that the coarse-grained DBSCAN cannot distinguish, and is used as the final point cluster classification result.

[0017] S1-2, wall column point cluster classification. Since there are significant differences in the geometric characteristics of columns and walls, specifically, columns have a smaller footprint and an aspect ratio close to 1, while walls usually have a larger footprint and a higher aspect ratio. Based on this, the main direction of each point cluster is calculated based on PCA, and the area and aspect ratio of each point cluster are calculated. If the area S of the point cluster is greater than the threshold S thr Or the aspect ratio AR is greater than the threshold AR thr , it is classified as a wall point cluster. Otherwise, PCA is used to determine whether its distribution is linear. If the point cluster distribution is not linear, it is classified as a column point cluster; if it is linear, it is classified as a short-line wall point cluster. For such short-line wall point clusters, the RANSAC algorithm can be directly applied to extract the corresponding straight line segments as the final analysis results.

[0018] S2. Extraction of column plane structure:

[0019] S2-1. Column point cluster denoising based on Gaussian kernel density estimation. Project the column point cluster onto a two-dimensional plane and perform grid division, and count the number of point clouds in each grid. i =[x i ,y i ] T And the number of point clouds in the grid w i As the weight, calculate the weighted mean μ and the corresponding weighted covariance matrix H, and the fitted Gaussian distribution is:

[0020]

[0021] Calculate the probability density value prob for each grid point i =f(g i ), and sort the grid points from high to low according to the probability density value; select the first k grid points by the cumulative probability density value, which satisfies the following formula:

[0022]

[0023] Where N kde is the preset cumulative probability density threshold. Finally, the original point cloud corresponding to these grid points is retained to achieve denoising.

[0024] S2-2. RANSAC-guided column structure extraction and optimization. Based on the denoised column point cloud and its pairwise perpendicular wall structure features, the RANSAC method is used to iteratively extract two mutually perpendicular straight line segments from the point cloud data, and then the corresponding rectangle is fitted as the planar structure of the current column point cluster.

[0025] S3. Wall plane structure extraction:

[0026] S3-1. Binarization and Skeleton Extraction of the Wall Point Cloud. The clustered wall point cloud data is projected onto the XOY plane, gridded, and a corresponding binary image is generated. Dilation and erosion operations are performed on the binary image. Based on this, the Zhang-Suen thinning algorithm is applied to extract the image's skeleton structure. By connecting the highlighted pixels in the skeleton image, the initial wall planar structure is obtained.

[0027] S3-2, Connected topology construction. Traverse all skeleton points in the skeleton image. For the i-th skeleton point Check the neighbor points in its 8-neighborhood (including up, down, left, right and diagonal directions), and only consider the connection if the neighbor point is a skeleton point. Then, prioritize connecting the neighbor skeleton points in the up, down, left and right directions; for the neighbor skeleton points in the diagonal direction, only if the current skeleton point does not exist in its 4-neighborhood (up, down, left and right) The connection is only made when there are other neighbor skeleton points.

[0028] S3-3, Feature Point Path Search. After the connected topology is constructed, the algorithm traverses all skeleton points and extracts endpoints (nodes with only one neighboring skeleton point) and trigeminals (nodes with three neighboring skeleton points) as feature points. If the entire connected topology is a loop (i.e., no endpoints or trigeminals), then corner points (with two neighboring skeleton points and the angle formed by the connection with other nodes is greater than or equal to the preset threshold) are extracted. ), and select the node closest to 90° from the corner points as the feature point. Then, the algorithm searches for the path between the feature points. The specific steps are as follows: Take a feature point vi As the starting point, expand the path along the connection relationship until the next feature point v is encountered j ; with v i , v j And all the nodes on the path are saved as a path. Then use v j As a new starting point, repeat the above process to search for paths between new feature points; during this period, duplicate paths are removed until all feature points are visited.

[0029] S3-4, straight line segment extraction. For each path obtained by the search, starting from the starting point v1 of the current path, gradually expand the next node v i Then, using all current nodes, we fit the straight line l by least squares. If each node v on the path i (i=1,...,n-1), the vertical distance to l is less than 0.5 pixel side length, then the current path is determined to be a straight path. Continue to expand the next node v i+1 .; If it does not meet the requirements, the newly expanded node v is discarded i , and the path {v1,v2,…,v i-1} as a new straight line path. Then use v i As the new starting point, we search for new straight line paths. After extracting all straight line paths, we generate corresponding straight line segments, and use the endpoints of these straight line segments to replace the original skeleton points to generate new connectivity relationships, thus obtaining the simplified planar structure of the wall point cluster.

[0030] S4, LiDAR and DXF point set cross-source registration:

[0031] S4-1, Feature Point Matching. Based on the initial scaling factor s0 provided by the DXF file, the column point set extracted from the DXF plane structure G is matched. Perform scale scaling to obtain the initial scaled point set Then, combined with the LiDAR column point set extracted in step S2 (i.e., the endpoint set that constitutes the plane structure of the LiDAR column), and generate the corresponding binary image I d with I l . Extract the corresponding ORB feature descriptor set D d With D l , feature coarse matching is achieved by minimizing the Hamming distance between descriptors.

[0032] S4-2, Feature point direction consistency screening. Statistical binary image I d with I lThe main direction differences of all matching point pairs in the histogram are calculated and a direction difference histogram is constructed. The matching point pairs within the first three largest intervals in the histogram are considered to be correct matching point pairs, while the matching point pairs falling in other intervals are considered to be incorrect matches and are eliminated.

[0033] S4-3, nearest neighbor consistency screening. Although the direction consistency check significantly reduces the number of incorrect matching point pairs, there may still be a small number of incorrect matches. To this end, the nearest neighbor consistency constraint is introduced to further improve the matching accuracy. In the specific implementation, for each matching point pair (D i ,D j )(where D i ∈D d , D j ∈D l ), respectively search for D in the remaining matching point pairs i .With D j The nearest neighbor descriptor of and If the following conditions are met:

[0034]

[0035] and is also a matching pair, then it is considered that (D i ,D j ) is the correct matching point pair. Where d(·,·) represents the Euclidean distance between corresponding points of the descriptor, T d is the preset distance threshold. Finally, the high-precision LiDAR matching point set is output. Matching point sets with DXF

[0036] S4-4, scaling factor estimation. For the point set P = {p1, p2, ..., p n}, first calculate its center of mass Then perform the centroid operation on each point to obtain p i ′=p i -c. Then, the scale factor s of the point set is expressed by the root mean square distance of the point set: Finally, the normalized point set is obtained In actual operation, the LiDAR matching point set Matching point sets with DXF Perform normalization operation to obtain the point set and point set and the corresponding scale factor s l With s d .based on and Calculate the covariance matrix H0:

[0037]

[0038] Perform SVD (Singular Value Decomposition) decomposition on the covariance matrix H0:

[0039] H0=U0Σ0V0 T (5)

[0040] Where U0 and V0 are orthogonal matrices, and Σ=diag(σ1,σ2) is a singular value matrix. According to the Helmert transformation, the scale factor between normalized point sets can be expressed as follows:

[0041]

[0042] Where, but Zoom to The scale factor required for the same scale is:

[0043]

[0044] S4-5, rigid body transformation calculation. Based on S1, DXF matching point set Perform scale scaling to obtain a point set Point Set Perform normalization again to obtain the point set Calculate the point set using formula (5) and The covariance matrix H1 between them is obtained, and H1 is decomposed by SVD using formula (6). Then, according to the Kabsch algorithm, Arrival Set The rigid body transformation is expressed as follows:

[0045]

[0046] Where V1 and U1 are orthogonal matrices obtained by SVD decomposition of H1; It is a point set The center of mass, c l It is a point set The center of mass.

[0047] Finally, based on the optimal scaling factor s fin =s0s1 and rigid body transformation T fin =(R fin ,t fin ), for the DXF plane structure point set P d Every point p in d,i Perform similarity transformation as shown below:

[0048]

[0049] Get the LiDAR point set P l Point sets that are precisely aligned in terms of scale characteristics and coordinate space distribution S5. Preliminary inspection of plane structure update:

[0050] S5-1, column structure update detection. First, calculate the column center point set extracted from the LiDAR point cloud The column center point set in DXF Where and m represent the number of columns extracted from LiDAR point cloud and the number of columns in DXF respectively. Based on these two point sets, the corresponding KD tree is constructed. and To speed up the efficiency of nearest neighbor search.

[0051] Column matching and classification: traversal The center point of each column use Query the nearest column center point in the DXF drawing And calculate the Euclidean distance d between the two ij According to d ij The comparison results with the preset thresholds thr1 and thr2 (satisfying thr1<thr2) classify the columns into the following three categories:

[0052] · Columns that have not been changed: If d ij ≤thr1, then it is considered The location has not changed.

[0053] New column: If d ij ≥thr2, then Marked as a newly added column.

[0054] Candidate matching pair: if thr1<d ij <thr2, then and marked as candidate column matching pairs

[0055] Such matching pairs may cause an overall offset in the extracted column position due to insufficient SLAM mapping accuracy, requiring subsequent pose correction and secondary update detection.

[0056] Further inspection of mismatched columns: The center point of the unmatched column use Query the nearest neighbor column center point in the LiDAR plane structure And calculate the Euclidean distance d between the two pq If dpq >thr1, then further check Is it close to a short linear wall? The specific steps are as follows:

[0057] 1. Traverse the plane structure corresponding to all wall clusters Filter out the wall clusters that contain only 2 vertices (i.e., a single straight line segment). Calculate the midpoints of these short linear walls and distance

[0058] 2. If It is believed that the short-line wall is due to data collection limitations or occlusion, resulting in the column point cloud being distributed in a short line and no corresponding column plane structure is generated. Mark as position unchanged. Then Mark the columns for deletion.

[0059] S5-2, wall structure update detection. The wall structure update detection is evaluated in units of wall clusters. Wall cluster validity evaluation: the wall plane structure G in DXF wall Each edge of the wall is interpolated and sampled at a fixed interval (0.1 meters) to generate a DXF wall point cloud. And build the corresponding KD tree To speed up the nearest neighbor search. For each wall cluster, the plane structure extracted in is the set of edges that make up the wall cluster. Interpolate and sample each edge at the same interval to generate the point cloud of the current wall cluster.

[0060] Traversal Each point in use Check its G w The nearest neighbor in And calculate the distance between the two like Then judge is a valid point. Calculate the current wall cluster Point cloud coverage, that is, the ratio of valid points to the total number of points i If the ratio i >ratio thr (90%), then the wall cluster is considered The edge (i.e. line segment) in G w If they exist, no update is required. Otherwise, mark The candidate wall clusters need to be updated for subsequent processing.

[0061] S6. Overall offset perception based on neighbor clustering:

[0062] S6-1. Clustering of candidate column matching points. For two candidate column center points p1 = (x1, y1) and p2 = (x2, y2), define the rectangular area between them as:

[0063] R(p1,p2)={(x,y)|x min <x<x max ,y min <y<y max} (10)

[0064] Where x min =min(x1,x2), x max =max(x1,x2),y min =min(y1,y2),y max = max(y1,y2). If there are no other column centers in R(p1,p2), then p1 and p2 are considered to be adjacent. Use the union-find data structure to merge adjacent points and output the clustering results. n c Indicates the number of clusters.

[0065] S6-2, adjacent column search. Based on the clustering result C of the candidate column matching points cols , further cluster the adjacent wall clusters. A candidate wall cluster in The sampled point cloud Construct the corresponding KD tree Traverse again The center point of each column use Query the sampling point cloud of the candidate cluster with the current wall The nearest neighbor point and solve the corresponding distance For all column center points According to distance Sort from small to large. Let the sequence of column center points after sorting be Where (i) represents the sorted index, satisfying is the center point of the i-th column after sorting and The shortest distance.

[0066] set up For the current cluster The adjacent column index set is initialized to Then, from the center point of the second column after sorting Start by determining the current column center To candidate wall clusters distance and The center of the last column in arrive distance Whether the relationship between the two meets the adjacency condition:

[0067]

[0068] If satisfied, then join in Otherwise terminate Where ∈ is the set distance threshold and needs to be smaller than the interval between columns.

[0069] S6-3, clustering of candidate wall clusters. After completing the adjacent column retrieval of each candidate wall cluster, further analyze the candidate cluster category of each column The association relationship between the candidate clusters and the wall clusters. Specifically, for each candidate cluster category If the current wall candidate cluster Adjacent columns exists in In the and Associated with The set of associated candidate wall clusters is denoted as n s Represents The number of associated wall candidate clusters. Thus, the corresponding clustering result pair can be constructed All clustering results are used To express.

[0070] S7. Structural optimization and secondary testing:

[0071] S7-1, structural optimization. Each clustering result in Perform ICP registration and solve a rigid body transformation (R s ,t s ), so that the distance between the corresponding points of the transformed candidate wall cluster point cloud and the column point and the target point cloud is minimized, thereby optimizing the position of the candidate column and the wall plane structure. Where (n) represents The pillar points extracted by LiDAR Matching DXF column point index, n s express The number of column matching point pairs in . As the source point cloud, its matching points are fixed and always right All candidate wall clusters in This paper uses the candidate wall cluster point cloud generated previously with interval as the interval Each point in Through KD tree Search DXF wall point clouds The nearest neighbor in as the corresponding matching points.

[0072] S7-2, Secondary test. Based on the optimal rigid body transformation obtained by ICP iteration, After optimizing the position of each clustering result pair, all candidate column matching points are further Candidate clusters with walls Perform a secondary test. Similarly, if Matching point If the distance is less than thr1, it is considered The position has not changed, otherwise, mark is the column that needs to be deleted, and It is marked as a new column. If its point cloud coverage is higher than the set threshold ratio thr , no update is required; otherwise, mark The wall cluster that needs to be updated for fusion.

[0073] S8. Update the column plan structure. Columns to be deleted are directly removed from the DXF plan structure G. For newly added columns, their coordinate information is first added to G. Then, the r nearest neighboring columns are retrieved from G. The average of these columns' geometric properties (for square columns, the size and orientation of the two orthogonal sides) is calculated and used as the size and orientation of the newly added column.

[0074] S9. Wall plane structure update:

[0075] S9-1, Initialization. Set the distance threshold D1, then traverse Unprocessed edges in If e 0 , if e 0 If the distance to the nearest edge of the DXF plane structure G is greater than or equal to D1, then e 0 For the newly added straight line segments, subsequent fusion tasks are required.

[0076] S9-2, Plane structure fusion based on depth-first search. From the selected edge e 0 Start along Perform a depth-first search on the edges in . For each traversed edge (u is the starting point, v is the end point), perform the following conditional judgment:

[0077] ●Condition 1: If the distance between vertex v and the nearest edge in G is greater than D2, then add the edge (u,v) to G and continue with v as the new starting point. Search for all edges (v, w) starting from v (i.e., the adjacent edges of v). Then repeat the depth-first search process for each adjacent edge.

[0078] Condition 2: If vertex v has not been visited and is close to the i-th edge in G Determine the edge (u,v) and the edge Is it close to vertical? If so, calculate the foot point v in Otherwise, find the point p closest to v, which is located on the edge of G

[0079] Up. Split at p, creating a new vertex v in , and add the edge (u,v in ) into G. At the same time, maintain a split vertex mapping table M and set M[v] = v in .

[0080] ●Condition 3: If vertex v has been visited, add the edge (u,M[v]) to G.

[0081] The final algorithm output is: all the updated results detected are integrated into the original plane structure G (DXF plane structure) according to the above update rules, and the updated plane structure is output. Among them, S1-S3 is the plane structure extraction stage; S4-S9 is the plane structure update stage.

[0082] Description of the accompanying tables and figures

[0083] Figure 1 This is a framework diagram of the method of the present invention.

[0084] Figure 2 This is the effect diagram of the point cloud clustering algorithm based on two-stage DBSCAN.

[0085] Figure 3 This is the rendering for distinguishing the wall column point clusters.

[0086] Figure 4 This is the denoising effect of the pillar point cloud based on Gaussian kernel density estimation.

[0087] Figure 5 Extract renderings for straight line segments based on path search.

[0088] Figure 6A comparison chart of the matching effects before and after ORB feature direction consistency screening.

[0089] Figure 7 Comparison of matching results before and after nearest neighbor consistency screening.

[0090] Figure 8 This is the cross-source registration effect diagram of LiDAR point set and DXF point set.

[0091] Figure 9 Schematic diagram of the planar structure of the overall offset columns and walls.

[0092] Figure 10 A diagram showing the SUPS scene dataset.

[0093] Figure 11 This is a data display diagram of the Quzhou B1 scene.

[0094] Figure 12 This is the DXF file of Quzhou B1 and the comparison of the wall point cloud change areas from April to July.

[0095] Figure 13 This is a data display diagram for Quzhou B2 scene.

[0096] Figure 14 Comparison of the plane structure extraction effects of various methods in different scenarios

[0097] Figure 15 The following is a comparison chart of the plane structure update effects of each method in different scenarios.

[0098] Figure 16 A comparison chart of the overall offset perception and calibration effects based on neighbor clustering. DETAILED DESCRIPTION

[0099] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments:

[0100] Step 1. Classification of wall and column point clusters. The wall structure of an underground parking lot mainly consists of two types: columns and walls. Columns are usually regular square cylinders with small spatial scales and consistent dimensions, and can be represented by a unified geometric model (such as a square or circle). In contrast, wall structures have significant diversity in layout form and spatial scale, making them difficult to describe with a single model. Given the significant differences in geometric characteristics and distribution patterns between walls and columns, accurately distinguishing between wall and column point clusters is a key prerequisite for extracting the planar structure of a parking lot.

[0101] Step 1-1. Point cloud clustering based on two-stage DBSCAN. The original wall point cloud is subjected to preliminary DBSCAN classification using coarse-grained parameter selection to obtain the initial clustering results. For each initial cluster C iFurther DBSCAN secondary classification is performed using fine-grained parameter selection. At the same time, the point cluster with the largest number of points in the newly generated point cluster is recorded. If the number of point clusters after secondary classification n is less than 2, that is, the fine-grained DBSCAN classification result is still a single cluster, then the initial point cluster C i As the final classification result. Otherwise, traverse all newly generated point clusters like The number of points in is less than the set threshold N fined , or its point cloud distribution shows linear characteristics (determined by PCA), then the point cluster is determined to be a small cluster generated by over-segmentation, and it is combined with the largest new point cluster Merge. On the contrary, if If the above conditions are not met, it is determined to be a valid point cluster that cannot be distinguished by coarse-grained DBSCAN and is used as the final point cluster classification result. Figure 2 As shown, Figure 2 (a), Figure 2 (b) with Figure 2 (c) Point cloud clustering effects of coarse-grained DBSCAN, fine-grained DBSCAN, and the proposed two-stage DBSCAN clustering algorithm.

[0102] Step 1-2. Classification of wall column point clusters. Since columns and walls have significant differences in geometric features, specifically, columns have a smaller footprint and an aspect ratio close to 1, while walls usually have a larger footprint and a higher aspect ratio. Based on this, PCA is used to calculate the main direction of each point cluster, and the area and aspect ratio of each point cluster are calculated. If the area S of the point cluster is greater than the threshold S thr Or the aspect ratio AR is greater than the threshold AR thr , then it is classified as a wall point cluster. Otherwise, PCA is further used to determine whether its distribution is linear. If the point cluster distribution is not linear, it is classified as a column point cluster; if it is linear, it is classified as a short-line wall point cluster. For such short-line wall point clusters, the RANSAC algorithm can be directly applied to extract the corresponding straight line segment as the final analysis result. The results are as follows Figure 3 As shown, Figure 3 (a), Figure 3 (b) The wall point clusters and column point clusters after differentiation, and different colors indicate different types of clusters.

[0103] Step 2. Extraction of column plane structure:

[0104] Step 2-1. Denoising of column point clusters based on Gaussian kernel density estimation. Project the column point clusters onto a two-dimensional plane and perform grid division, and count the number of point clouds in each grid. Based on the grid center position g i =[x i ,y i ]T And the number of point clouds in the grid w i As the weight, calculate the weighted mean μ and the corresponding weighted covariance matrix H, and the fitted Gaussian distribution is:

[0105]

[0106] Calculate the probability density value prob for each grid point i =f(g i ), and sort the grid points from high to low according to the probability density value; select the first k grid points by the cumulative probability density value, which satisfies the following formula:

[0107]

[0108] Where N kde is the preset cumulative probability density threshold. Finally, the original point cloud data corresponding to these grid points are retained to achieve denoising. Figure 4 The comparison of the column point cluster effects before and after denoising is shown, where (a) and (c) are the column point clusters before denoising, and (b) and (d) are the results after denoising.

[0109] Step 2-2. RANSAC-guided column structure extraction and optimization: Based on the denoised column point cloud and its pairwise perpendicular wall structure features, the RANSAC method is used to iteratively extract two mutually perpendicular straight line segments from the point cloud data, and the corresponding rectangles are further fitted as the planar structure of the current column point cluster.

[0110] Step 3. Wall plane structure extraction:

[0111] Step 3-1. Binarization of the wall point cloud and skeleton extraction. The clustered wall point cloud data is projected onto the XOY plane, gridded, and a corresponding binary image is generated. Dilation and erosion operations are performed on the binary image. Based on this, the Zhang-Suen thinning algorithm is applied to extract the image's skeleton structure. By connecting the highlighted pixels in the skeleton image, the initial wall planar structure is obtained.

[0112] Step 3-2. Connected topology construction. Traverse all skeleton points in the skeleton image. For the i-th skeleton point Check the neighbor points in its 8-neighborhood (including up, down, left, right and diagonal directions), and only consider the connection if the neighbor point is a skeleton point. Then, prioritize connecting the neighbor skeleton points in the up, down, left and right directions; for the neighbor skeleton points in the diagonal direction, only if the current skeleton point does not exist in its 4-neighborhood (up, down, left and right) The connection is only made when there are other neighbor skeleton points.

[0113] Step 3-3. Feature point path search. After the connected topology is constructed, the algorithm traverses all skeleton points and extracts endpoints (nodes with only one neighboring skeleton point) and trigeminals (nodes with three neighboring skeleton points) as feature points. If the entire connected topology is a loop (i.e., no endpoints or trigeminals), then corner points (with two neighboring skeleton points and the angle formed by connecting with other nodes is greater than or equal to the preset threshold) are extracted. ), and select the node closest to 90° from the corner points as the feature point. Then, the algorithm searches for the path between the feature points. The specific steps are as follows: Take a feature point v i As the starting point, expand the path along the connection relationship until the next feature point v is encountered j ; with v i , v j And all the nodes on the path are saved as a path. Then use v j As a new starting point, repeat the above process to search for paths between new feature points; during this period, duplicate paths are removed until all feature points are visited.

[0114] Step 3-4. Line segment extraction. For each path found, start from the starting point v1 of the current path and gradually expand to the next node v i Then, using all current nodes, we fit the straight line l by least squares. If each node v on the path i (i=1,...,n-1), the vertical distance to l is less than 0.5 pixel side length, then the current path is determined to be a straight path. Continue to expand the next node v i+1 .; If it does not meet the requirements, the newly expanded node v is discarded i , and the path {v1,v2,...,v i-1} as a new straight line path. Then use v i As the new starting point, we search for new straight line paths. After extracting all straight line paths, we generate corresponding straight line segments, and use the endpoints of these straight line segments to replace the original skeleton points to generate new connectivity relationships, thus obtaining the simplified planar structure of the wall point cluster. Figure 5 The algorithm effect is demonstrated, verifying that it can preserve the sharp features of the wall corners while simplifying the plane structure.

[0115] Step 4. Cross-source registration of LiDAR and DXF point sets:

[0116] Step 4-1. Feature point matching. Based on the initial scaling factor s0 provided by the DXF file, the column point set extracted from the DXF plane structure G is matched. Perform scale scaling to obtain the initial scaled point set Then, combined with the LiDAR column point set extracted in step S2 (i.e., the endpoint set that constitutes the plane structure of the LiDAR column), and generate the corresponding binary image I d with I l . Extract the corresponding ORB feature descriptor set D d With D l , feature coarse matching is achieved by minimizing the Hamming distance between descriptors.

[0117] Step 4-2. Feature point direction consistency screening. Statistical binary image I d with I l The main direction differences of all matching point pairs in the histogram are calculated and a direction difference histogram is constructed. The matching point pairs within the first three largest intervals in the histogram are considered to be correct matching point pairs, while the matching point pairs falling in other intervals are considered to be incorrect matches and are eliminated. Figure 6 The comparison of matching effects before and after ORB feature direction consistency screening is shown.

[0118] Step 4-3. Nearest neighbor consistency screening. Although the direction consistency check significantly reduces the number of incorrect matching point pairs, there may still be a small number of incorrect matches. To this end, the nearest neighbor consistency constraint is introduced to further improve the matching accuracy. In the specific implementation, for each matching point pair (D i ,D j )(where D i ∈D d , D j ∈D l ), respectively search for D in the remaining matching point pairs i .With D j The nearest neighbor descriptor of and If the following conditions are met:

[0119] and is also a matching pair, then it is considered that (D i ,D j ) is the correct matching point pair. Where d(·,·) represents the Euclidean distance between corresponding points of the descriptor, T d is the preset distance threshold. Finally, the high-precision LiDAR matching point set is output. Matching point sets with DXF Figure 7 The comparison of matching effects before and after nearest neighbor consistency screening is shown.

[0120] Step 4-4. Scaling factor estimation. For the point set P = {p1, p2, ..., p n}, first calculate its center of mass Then perform the centroid operation on each point to obtain p i′=p i -c. Then, the scale factor s of the point set is expressed by the root mean square distance of the point set: Finally, the normalized point set is obtained In actual operation, the LiDAR matching point set Matching point sets with DXF Perform normalization operation to obtain the point set and point set and the corresponding scale factor s l With s d .based on and Calculate the covariance matrix H0:

[0121]

[0122] Perform SVD (Singular Value Decomposition) decomposition on the covariance matrix H0:

[0123] H0=U0Σ0V0 T (16)

[0124] Where U0 and V0 are orthogonal matrices, and Σ=diag(σ1,σ2) is a singular value matrix. According to the Helmert transformation, the scale factor between normalized point sets can be expressed as follows:

[0125]

[0126] Where, but Zoom to The scale factor required for the same scale is:

[0127]

[0128] Step 4-5. Calculate the rigid body transformation. Based on s1, calculate the DXF matching point set. Perform scale scaling to obtain a point set Point Set Perform normalization again to obtain the point set Calculate the point set using formula (16) and The covariance matrix H1 between them is obtained, and H1 is decomposed by SVD using formula (17). Then, according to the Kabsch algorithm, Arrival Set The rigid body transformation is expressed as follows:

[0129]

[0130] Where V1 and U1 are orthogonal matrices obtained by SVD decomposition of H1; with c l Point sets and The center of mass.

[0131] Finally, based on the optimal scaling factor s fin =s0s1 and rigid body transformation T fin =(R fin ,t fin ), for the DXF plane structure point set P d Every point p in d,i Perform similarity transformation as shown below:

[0132]

[0133] Get the LiDAR point set P l Point sets that are precisely aligned in terms of scale characteristics and coordinate space distribution Figure 8 The cross-source registration effect of LiDAR point set and DXF point set is demonstrated.

[0134] Step 5. Preliminary inspection of plane structure update:

[0135] Step 5-1. Update detection of column structure. First, calculate the column center point set extracted from the LiDAR point cloud. The column center point set in DXF Where and m represent the number of columns extracted from LiDAR point cloud and the number of columns in DXF respectively. Based on these two point sets, the corresponding KD tree is constructed. and To speed up the efficiency of nearest neighbor search.

[0136] Column matching and classification: traversal The center point of each column use Query the nearest column center point in the DXF drawing And calculate the Euclidean distance d between the two ij According to d ij The comparison results with the preset thresholds thr1 and thr2 (satisfying thr1<thr2) classify the columns into the following three categories:

[0137] ●Unchanged columns: If d ij ≤thr1, then it is considered The location has not changed.

[0138] ●New column: If d ij ≥thr2, then Marked as a newly added column.

[0139] ●Candidate matching pair: if thr1<d ij <thr2, then and marked as candidate column matching pairs

[0140] Such matching pairs may cause an overall offset in the extracted column position due to insufficient SLAM mapping accuracy, requiring subsequent pose correction and secondary update detection.

[0141] Further inspection of mismatched columns: The center point of the unmatched column use Query the nearest neighbor column center point in the LiDAR plane structure And calculate the Euclidean distance d between the two pq If d pq >thr1, then further check Is it close to a short linear wall? The specific steps are as follows:

[0142] 1. Traverse the plane structure corresponding to all wall clusters Filter out the wall clusters that contain only 2 vertices (i.e., a single straight line segment). Calculate the midpoints of these short linear walls and distance

[0143] 2. If It is believed that the short-line wall is due to data collection limitations or occlusion, resulting in the column point cloud being distributed in a short line and no corresponding column plane structure is generated. Mark as position unchanged. Then Mark the columns for deletion.

[0144] Step 5-2. Wall structure update detection: The wall structure update detection is evaluated in units of wall clusters.

[0145] Wall cluster validity evaluation: wall plane structure G in DXF wall Each edge of the wall is interpolated and sampled at a fixed interval (0.1 meters) to generate a DXF wall point cloud. And build the corresponding KD tree To speed up the nearest neighbor search. For each wall cluster, the plane structure extracted in is the set of edges that make up the wall cluster. Interpolate and sample each edge at the same interval to generate the point cloud of the current wall cluster. Traversal Each point in use Check its G w The nearest neighbor in And calculate the distance between the two like Then judge is a valid point. Calculate the current wall cluster Point cloud coverage, that is, the ratio of valid points to the total number of points i If the ratio i >ratio thr (90%), then the wall cluster is considered The edge (i.e. line segment) in G w If they exist, no update is required. Otherwise, mark The candidate wall clusters need to be updated for subsequent processing.

[0146] Step 6. Global offset perception algorithm based on neighbor clustering:

[0147] This algorithm is proposed to solve the problem of overall offset of wall point cloud caused by insufficient SLAM mapping accuracy, which in turn causes misjudgment of plane structure update. Figure 9 As shown in the figure, the red line represents the plane structure extracted based on the LiDAR point cloud, and the blue line represents the plane structure extracted from the DXF file. Figure 9 The local plane structure extraction effect shown in the orange circle in the middle shows that although the extracted plane structure is highly similar to the true value in terms of morphological distribution, there is still a significant overall pose deviation. This deviation is consistent with the column offset problem and is mainly attributed to the insufficient accuracy of SLAM map construction. For example, in the corner area, the rapid rotation of the sensor causes a significant increase in the pose estimation error, which in turn causes the overall position offset of the wall point cloud, ultimately affecting and causing misjudgment of the plane structure update. The process of the overall offset perception algorithm based on adjacency clustering is as follows:

[0148] Step 6-1. Clustering of candidate column matching points. For two candidate column center points p1 = (x1, y1) and p2 = (x2, y2), define the rectangular area between them as:

[0149] R(p1,p2)={(x,y)|x min <x<x max ,y min <y<y max} (twenty one)

[0150] Where x min =min(x1,x2), x max =max(x1,x2),y min =min(y1,y2),ymax = max(y1,y2). If there are no other column centers in R(p1,p2), then p1 and p2 are considered to be adjacent. Use the union-find data structure to merge adjacent points and output the clustering results. n c Indicates the number of clusters.

[0151] Step 6-2. Adjacent pillar search. Based on the clustering result C of candidate pillar matching points cols , further cluster the adjacent wall clusters. A candidate wall cluster in The sampled point cloud Construct the corresponding KD tree Traverse again The center point of each column use Query the sampling point cloud of the candidate cluster with the current wall The nearest neighbor point and solve the corresponding distance For all column center points According to distance Sort from small to large. Let the sequence of column center points after sorting be Where (i) represents the sorted index, satisfying is the center point of the i-th column after sorting and The shortest distance.

[0152] set up For the current cluster The adjacent column index set is initialized to Then, from the center point of the second column after sorting Start by determining the current column center To candidate wall clusters distance and The center of the last column in arrive distance Whether the relationship between the two meets the adjacency condition:

[0153]

[0154] If satisfied, then join in Otherwise terminate Where ∈ is the set distance threshold and needs to be smaller than the interval between columns.

[0155] Step 6-3. Clustering of candidate wall clusters. After completing the adjacent column retrieval of each candidate wall cluster, further analyze the candidate cluster category of each column The association relationship between the candidate clusters and the wall clusters. Specifically, for each candidate cluster category If the current wall candidate cluster Adjacent columns exists in In the and Associated with The set of associated candidate wall clusters is denoted as n s Represents The number of associated wall candidate clusters. Thus, the corresponding clustering result pair can be constructed All clustering results are used To express.

[0156] Step 7. Structural optimization and secondary testing:

[0157] Step 7-1. Structural optimization. Each clustering result in Perform ICP registration and solve a rigid body transformation (R s ,t s ), so that the distance between the corresponding points of the transformed candidate wall cluster point cloud and the column point and the target point cloud is minimized, thereby optimizing the position of the candidate column and the wall plane structure. Where (n) represents The pillar points extracted by LiDAR Matching DXF column point index, n s express The number of column matching point pairs in . As the source point cloud, its matching points are fixed and always right All candidate wall clusters in This paper uses the candidate wall cluster point cloud generated previously with interval as the interval Each point in Through KD tree Search DXF wall point clouds The nearest neighbor in as the corresponding matching points.

[0158] Step 7-2. Secondary test. Based on the optimal rigid body transformation obtained by ICP iteration, After optimizing the position of each clustering result pair, all candidate column matching points are further Candidate clusters with walls Perform a secondary test. Similarly, if Matching point If the distance is less than thr1, it is considered The position has not changed, otherwise, mark is the column that needs to be deleted, and It is marked as a new column. If its point cloud coverage is higher than the set threshold ratio thr , no update is required; otherwise, mark The wall cluster that needs to be updated for fusion.

[0159] Step 8. Update the column plan structure. Columns to be deleted are directly removed from the DXF plan structure G. For newly added columns, their coordinate information is first added to G. Then, G retrieves the r nearest neighboring columns and calculates the average of their geometric properties (for square columns, the size and orientation of the two orthogonal sides). This average is used as the size and orientation of the newly added column.

[0160] Step 9. Wall plane structure update:

[0161] Step 9-1. Initialization. Set the distance threshold D1 and then traverse Unprocessed edges in If e 0 , if e 0 If the distance to the nearest edge of the DXF plane structure G is greater than or equal to D1, then e 0 For the newly added straight line segments, subsequent fusion tasks are required.

[0162] Step 9-2. Planar structure fusion based on depth-first search. From the selected edge e 0 Start along Perform a depth-first search on the edges in . For each traversed edge (u is the starting point, v is the end point), perform the following conditional judgment:

[0163] Condition 1: If the distance between vertex v and the nearest edge in G is greater than D2, then add the edge (u, v) to G and continue with v as the new starting point. Search for all edges (v, w) starting from v (i.e., the adjacent edges of v). Then repeat the depth-first search process for each adjacent edge.

[0164] Condition 2: If vertex v has not been visited and is close to the i-th edge in G Determine the edge (u,v) and the edge Is it close to vertical? If so, calculate the foot point v in Otherwise, find the point p closest to v, which is located on the edge of G Up. Split at p, creating a new vertex v in , and add the edge (u,v in ) into G. At the same time, maintain a split vertex mapping table M and set M[v] = v in .

[0165] Condition 3: If vertex v has been visited, add the edge (u, M[v]) to G.

[0166] The final algorithm output is: all the updated results detected are integrated into the original plane structure G (DXF plane structure) according to the above update rules, and the updated plane structure is output. Among them, steps 1 to 3 are the plane structure extraction stage; steps 4 to 9 are the plane structure update stage.

[0167] Example

[0168] The data sets of this experiment are derived from the public underground parking simulation data set SUPS and the actual data collected from the underground first floor (B1) and underground second floor (B2) of the parking lot of the Quzhou Research Institute of the University of Electronic Science and Technology of China. Figure 10-13 The simulation model of the SUPS scenario is shown in Figure 10 (a) shows the use of LIO-SAM to construct a parking lot lidar point cloud map. Figure 10 (b) Figure 11 (b) Figure 13 (b) As shown. After a series of preprocessing operations, the corresponding laser radar wall point cloud maps are obtained, as shown in Figure 10 (c), Figure 11 (c), Figure 13 (c) shown.

[0169] Regarding the acquisition of true value data in the plane structure extraction evaluation experiment: For the SUPS scene, in the professional point cloud processing software CloudCompare, the true value file is generated by manually annotating the segmented wall point cloud; for the Quzhou B1 and Quzhou B2 scenes, the DXF files drawn by professional surveying and mapping practitioners are parsed into OSM files as the true value data. The original DXF files of the parking lot in the Quzhou B1 and Quzhou B2 scenes are as follows: Figure 11 (a) with Figure 13 (a)

[0170] For the experimental evaluation of planar structure updating, the SUPS and Quzhou B1 scenes were used to evaluate the actual effectiveness of the updating method. The following focuses on the detailed description of the construction process of the SUPS and Quzhou B1 datasets.

[0171] SUPS dataset: (1) Existing plane structure generation: Based on the wall point cloud of the original SUPS scene, the professional point cloud processing software CloudCompare is used to manually draw the surface structure to be updated. (2) Generation of point cloud after change: The distribution of the original wall point cloud is modified, and the wall point cloud after change is manually drawn as the experimental input, and the corresponding ground truth plane structure is generated at the same time. The wall point cloud of the changed SUPS scene is as follows: Figure 10 (d) shown.

[0172] Quzhou B1 dataset: (1) Existing plane structure: The DXF file of Quzhou B1 is used as the original plane structure. (2) Acquisition of true plane structure: The point cloud maps of the Quzhou B1 scene were collected in April and July 2024, and their changes were compared and analyzed. The wall point cloud of the Quzhou B1 scene after the change is as follows: Figure 11 (d) For the convenience of display and comparison, Figure 12 (a), Figure 12 (b) and Figure 12 (c) Comparison of the change regions of the DXF file and the wall point cloud data for Quzhou in April and July. The wall point cloud collected in July is used as the actual distribution, and the corresponding wall plane structure is added to the original DXF file as the ground truth for the experiment.

[0173] This paper designs multiple sets of experiments to verify the superiority of the proposed method. The first set of experiments compares the proposed planar structure extraction method with some existing indoor planar structure extraction methods in various scenarios. The main evaluation indicators are the precision, recall rate and F1 score of the line segment extraction in the extracted structure. [3] The experimental methods for comparison are as follows: (1) IRC [2] :This method detects short line segments based on the iterative RANSAC clustering algorithm, and restores the topological relationship between line segments through post-processing steps such as line segment direction optimization, fusion and connection, and finally generates a planar structure. (2) FGNP [3] : This method targets high-noise cloud scenes and uses the PCA linearly constrained region growing algorithm to extract initial short line segments. It then optimizes the topological relationship by deleting redundant line segments, merging adjacent line segments, and connecting breakpoints to reconstruct the planar structure.

[0174] Table 1 Performance comparison of various methods in different scenarios

[0175]

[0176] Table 1 shows the performance comparison of various methods in line segment extraction tasks under different scenarios, where bold indicates the best result, and Ours indicates the plane structure extraction method proposed in this patent. Experimental results show that plane structure extraction methods such as IRC and FGNP, ​​which are based on short line segment detection and fusion, have a strong dependence on the quality of the input point cloud. In a wide range of actual scenarios, the thickness of the wall point cloud increases due to the cumulative error of SLAM mapping, which significantly reduces the accuracy of short line segment extraction, thereby affecting the topological structure reconstruction effect, resulting in poor performance of these two methods in line segment extraction accuracy (especially the accuracy rate indicator). In contrast, the plane structure extraction method proposed in this patent achieved the best results in all scenarios.

[0177] The second set of experiments qualitatively compared the proposed planar structure extraction method with the extraction results of IRC and FGNP and the visual effects of their local details in various scenarios. Figure 14 As shown in the figure, GT represents the true value of the plane structure. It can be seen that when dealing with wall point clouds with significant thickness and noise interference, the short line segment extraction accuracy (especially direction estimation) of strategies based on short line segment detection and fusion such as IRC and FGNP is significantly reduced, resulting in poor results in subsequent line segment fusion and topology reconstruction. The method proposed in this patent introduces a straight line segment extraction algorithm based on path search on the basis of morphological skeleton extraction. While ensuring the accuracy of plane structure extraction, it significantly improves the smoothness and aesthetics of the extraction results, fully verifying the robustness and superiority of the proposed plane structure extraction method in different scenarios.

[0178] The third set of experiments qualitatively compared the visual effects of the proposed planar structure update method with the existing indoor planar structure update method in multiple scenarios with wall planar structure changes. The experimental methods for comparison are: CPSEFuse: The planar structure extraction part of this method adopts the planar structure update method proposed in this patent and is based on Bastani et al. [5] The proposed fusion algorithm extends the plane structure extraction method to the plane structure update task. The fusion algorithm achieves the plane structure update by fusing the extraction results with the existing plane structure. Figure 15 As shown in the figure, Old represents the original plane structure and GT represents the updated plane structure truth value. It can be seen that the plane structure update method proposed in this patent can effectively identify and remove the column structures that exist in the original structure but not in the truth structure, while achieving accurate and complete fusion of the newly added wall and column structures. In contrast, due to the fusion strategy adopted by CPSEFuse, it only achieves the detection and fusion of the plane structure of the newly added wall, but cannot remove the column structures that need to be deleted.

[0179] The fourth set of experiments verified the effectiveness of the overall offset perception method proposed in this patent. The results are as follows: Figure 16 shown. Figure 16 (a) with Figure 16 (b) shows the problem of position offset of the wall plane structure caused by insufficient SLAM mapping accuracy and the effects before and after correction. The red line represents the extracted plane structure, and the blue line represents the true plane structure. The experimental results show that the overall offset perception method based on adjacency clustering and the subsequent calibration mechanism proposed in this paper can effectively optimize the position of the offset wall plane structure, making it highly consistent with the true plane structure. This optimization process successfully avoids the misjudgment of the plane structure update caused by position offset, and ensures the consistency of the final update result with the original plane structure and the true value structure, as shown in Figure 2. Figure 16 (c) In contrast, the update strategy of the comparative experiment failed to identify and handle the overall position offset problem caused by insufficient SLAM accuracy, resulting in the offset structure being incorrectly added to the original planar structure, as shown in Figure 16 As shown in (d), it can be seen that the proposed method significantly improves the robustness and accuracy of planar structure update through precise offset perception and calibration.

Claims

1. A clustered point cloud-driven underground parking lot planar structure extraction and updating method, characterized by: The workflow of the method includes two stages: plane structure extraction and plane structure update. In the plane structure extraction stage, to address the problem of wall structure heterogeneity in underground parking environments, the method uses a two-stage DBSCAN algorithm to accurately cluster point cloud data; in addition, the method introduces a column point cluster denoising algorithm based on Gaussian kernel density estimation and a straight line segment extraction algorithm based on path search to solve the problem of decreased plane structure extraction accuracy caused by the degradation of wall point cloud quality in underground parking environments. In the plane structure update stage, the method implements cross-source registration of LiDAR and DXF through feature consistency constraints. To address the problem of overall offset of wall point clouds caused by insufficient SLAM mapping accuracy, the overall offset perception algorithm based on adjacency clustering is used to identify and calibrate the offset structure, thereby avoiding misjudgment in update detection.

2. The two-stage DBSCAN algorithm according to claim 1, characterized in that: The original wall point cloud is subjected to preliminary DBSCAN classification using coarse-grained parameter selection to obtain the initial clustering results. Each initial cluster is further subjected to secondary DBSCAN classification using fine-grained parameter selection. At the same time, the main cluster with the largest number of points in the newly generated point clusters is recorded. If the number of point clusters after secondary classification is less than 2, the initial clustering result is retained. Otherwise, all newly generated point clusters are traversed. If the number of points in the cluster is lower than the set threshold, or its point cloud distribution shows linear characteristics (determined by PCA), the point cluster is determined to be a small point cluster generated by over-segmentation and is merged with the main cluster. Otherwise, it is determined to be a valid point cluster that the coarse-grained DBSCAN cannot distinguish, and is used as the final point cluster classification result.

3. The column point cluster denoising algorithm based on Gaussian kernel density estimation according to claim 1, characterized in that: The column point cluster is projected onto a two-dimensional plane and meshed, with the number of point clouds within each grid counted. A weighted covariance matrix is ​​constructed based on the grid center position and the number of point clouds within the grid as weights. Subsequently, the probability density value of each grid cell is calculated using Gaussian kernel density estimation, and the grid points are sorted in descending order by probability density value. By accumulating the probability density values, all wall grids whose cumulative probability density value falls within a pre-set confidence interval are retained, thus achieving denoising of the column point cluster.

4. The straight line segment extraction algorithm based on path search according to claim 1, characterized in that: Based on the morphological skeleton extraction results, a connectivity topology of the skeleton points is established according to the preset connectivity rules. Subsequently, the feature points in the topology are identified and all shortest paths between the feature points are searched. Finally, for each searched path, the path nodes are gradually expanded from the starting point and the corresponding straight line paths are extracted. After all the straight line paths are extracted, the corresponding straight line segments are generated. The endpoints of these straight line segments are used to replace the original skeleton points to generate new connectivity relationships, thereby obtaining the simplified planar structure of the wall point cluster.

5. The overall offset perception algorithm based on adjacency clustering according to claim 1, characterized in that: First, adjacent candidate column matching points are identified based on adjacency relationships and clustered using a union-find data structure. Subsequently, based on the clustering results of the candidate column matching points, each candidate wall cluster is traversed to retrieve its spatially adjacent columns. Finally, by analyzing the geometric relationship between each column candidate cluster and the candidate wall cluster, if the adjacent columns of a candidate wall cluster belong to a specific column cluster, the two are considered associated, and a joint cluster pair is constructed for the corresponding candidate wall cluster and candidate column cluster. Based on this joint cluster pair, all point clusters within it are inferred to follow the same spatial offset pattern, thus achieving holistic offset perception of planar structures.

Images