A Method and System for Automatic Parking Map Generation Based on Point Cloud and Semantic Features

Abstract

This invention provides a method and system for automatically generating parking lot maps based on point cloud and semantic features, belonging to the field of parking lot map construction technology. The method includes: acquiring the original LiDAR point cloud of the parking lot and preprocessing it to obtain a ground point cloud; extracting parking space line point clouds based on the reflection intensity of each point from the obtained ground point cloud; clustering and segmenting the parking space line point clouds to obtain multiple independent parking space line point cloud clusters; performing straight line fitting on each parking space line point cloud cluster to obtain parking space boundary lines and parking space dividing lines; generating parking space boxes based on the intersections of the parking space boundary lines and parking space dividing lines, and outputting the parking lot map. This invention uses a reflection intensity threshold segmentation method to extract parking space line point clouds, overcoming the influence of illumination changes on parking space line extraction; by combining density clustering and a two-step straight line fitting strategy based on geometric constraints, it effectively solves the problems of ineffective differentiation of continuous parking space boundaries and insufficient fitting accuracy of dividing lines.

Description

Technical Field

[0001] This invention relates to the field of parking lot map construction technology, and in particular to a method and system for automatically generating parking lot maps based on point clouds and semantic features. Background Technology

[0002] Parking lot maps are the core data support for unmanned parking, intelligent scheduling, and dynamic monitoring of parking spaces, and their level of automation in generation directly affects the precision of parking lot management. However, the current generation of parking lot maps mainly relies on manual drawing and annotation, which leads to problems such as long annotation cycles, high labor costs, and delayed updates when dealing with large-scale parking lots, making it difficult to meet the real-time map requirements of smart parking lots.

[0003] To address the aforementioned issues, existing research primarily employs computer vision methods, using cameras to capture images and deep learning techniques to identify parking space lines, thereby generating parking lot maps. However, visual images are susceptible to changes in lighting and weather conditions, and issues such as lens distortion and obstructions can directly compromise the integrity of parking space lines, leading to poor detection stability and a high risk of failure. Furthermore, these methods rely on large amounts of labeled data and computational resources, making practical deployment challenging.

[0004] In recent years, LiDAR technology has been widely used in environmental perception and spatial modeling due to its advantages such as high precision and strong resistance to light interference. Some commercial products have achieved automatic construction of 3D models based on LiDAR, but the generated maps are only 3D mesh models and cannot achieve automatic recognition and labeling of parking space lines. There is still a technological gap in the generation of parking space-level maps.

[0005] Therefore, automatically generating parking lot maps from laser point clouds faces two major technical challenges: first, how to accurately extract parking space line point clouds in scenes with varying lighting and darkness, avoiding point cloud incompleteness and misidentification caused by uneven reflection intensity or occlusion; second, how to effectively segment and fit continuous parking space point clouds into straight lines, avoiding conventional methods that fit all parking spaces into a straight line and fail to distinguish the boundaries of individual parking spaces. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for automatically generating parking lot maps based on point cloud and semantic features, so as to solve the technical problems in the prior art where parking space line extraction is easily affected by lighting interference and continuous parking spaces cannot be effectively distinguished, and to realize the automatic generation of parking lot maps.

[0007] To achieve the above objectives, this invention provides a method for automatically generating parking lot maps based on point clouds and semantic features, comprising the following steps: Step S1: Obtain the original LiDAR point cloud of the parking lot and preprocess the original LiDAR point cloud to obtain the ground point cloud; Step S2: Based on the obtained ground point cloud, extract the parking space line point cloud according to the reflection intensity of each point, and perform clustering and segmentation on the parking space line point cloud to obtain multiple independent parking space line point cloud clusters. Step S3: Perform straight line fitting on each parking space line point cloud cluster to obtain the parking space boundary line and parking space dividing line; Step S4: Generate parking space frames based on the intersection of parking space boundary lines and parking space separator lines, and output the parking lot map.

[0008] Preferably, in step S1, the preprocessing includes: A statistical outlier removal filtering algorithm is used to denoise the original lidar point cloud in order to remove outlier noise points; A slope-based ground filtering algorithm is used to separate ground points from non-ground points in the denoised point cloud to obtain the ground point cloud.

[0009] In a preferred slope-based ground filtering algorithm, the maximum allowable height difference between two points is... The calculation is as follows: ; in, The maximum allowable slope angle is set. The horizontal distance between the two points This is for height offset tolerance.

[0010] Preferably, in step S2, extracting the parking space line point cloud based on the reflection intensity of each point specifically includes: Statistically calculate the mean reflection intensity at each point in the ground point cloud. and standard deviation ; The reflection intensity threshold is determined based on the mean and standard deviation. : ; in, It is a multiple of the standard deviation; The reflection intensity is greater than the reflection intensity threshold. The point cloud was extracted as parking space line point cloud.

[0011] Preferably, in step S2, the density-based clustering algorithm DBSCAN is used to cluster and segment the parking space line point cloud, dividing the spatially continuous parking space line point cloud into the same parking space line point cloud cluster.

[0012] Preferably, in step S3, the linear fitting of each parking space line point cluster specifically includes: Step S31: Use the random sampling consensus algorithm to perform straight line fitting on the current parking space line point cloud cluster to obtain the main straight line, which serves as the parking space boundary line. Step S32: Remove the inner points corresponding to the main line from the current parking space line point cloud cluster, and repeatedly perform the random sampling consensus algorithm to fit candidate lines on the remaining point cloud; Step S33: Calculate the angle between the candidate line and the main line. If the angle is less than the angle threshold, the candidate line is retained as another parking space boundary line and its slope is adjusted to be the same as the main line. If the angle is greater than or equal to the angle threshold, the candidate line is discarded.

[0013] Preferably, in step S3, the linear fitting of each parking space line point cluster further includes: After fitting all parking space boundary lines, density-based clustering algorithm is used again to finely cluster the dividing line point cloud from the remaining point cloud of the current parking space line point cloud cluster, resulting in multiple dividing line point cloud sub-clusters. For each sub-cluster of the dividing line point cloud, a straight line is fitted using a random sampling consensus algorithm to obtain the parking space dividing line, and the slope of the parking space dividing line is forced to satisfy the vertical geometric constraint with the slope of the parking space boundary line.

[0014] Preferably, in step S4, generating the parking space frame specifically includes: Calculate the four intersection points of the two adjacent parking space separator lines and the two parking space boundary lines; Connect the four intersection points in spatial order to form a closed rectangle, which serves as the parking space frame.

[0015] Preferably, in step S4, the output parking lot map contains a unique number for each parking space frame.

[0016] This invention also provides an automatic parking map generation system based on point clouds and semantic features, for performing the automatic parking map generation method based on point clouds and semantic features as described above, including: The point cloud acquisition and preprocessing module is used to acquire the original LiDAR point cloud of the parking lot and preprocess the original LiDAR point cloud to obtain the ground point cloud. The parking line extraction and clustering module is used to extract parking line point clouds based on ground point clouds and to cluster and segment the parking line point clouds to obtain multiple independent parking line point cloud clusters. The line fitting module is used to perform line fitting on each parking space line point cloud cluster to obtain the parking space boundary line and parking space dividing line. The map generation module is used to generate parking space frames based on the intersection of parking space boundary lines and parking space separator lines, and output a parking lot map.

[0017] Therefore, the present invention employs the above-mentioned method and system for automatically generating parking lot maps based on point clouds and semantic features, and the beneficial technical effects are as follows: (1) This invention uses the reflection intensity threshold segmentation method to extract parking space line point clouds. It utilizes the physical characteristic that the reflection intensity of lidar is determined by the target material and is not affected by ambient light, thus overcoming the technical defects of traditional computer vision methods that rely on RGB color information and are easily affected by changes in lighting. It can stably extract complete parking space line point clouds under different lighting conditions, achieving all-weather robust detection performance.

[0018] (2) This invention employs a two-step clustering and geometric constraint-based straight-line fitting strategy: First, continuous parking spaces are divided into independent clusters using DBSCAN coarse clustering. For each cluster, the RANSAC algorithm is used to fit the parking space boundary line, and parallel constraints are introduced to filter effective long sides. After removing the long-side point cloud, the remaining point cloud is finely clustered and a dividing line is fitted. Vertical constraints are introduced to ensure that the short side is strictly perpendicular to the long side. This strategy overcomes the shortcomings of conventional straight-line fitting methods that fit all continuous parking spaces into a straight line and cannot distinguish the boundaries of individual parking spaces, thus achieving accurate identification in continuous parking space scenarios.

[0019] (3) The present invention adopts a method of generating parking space frames based on the intersection of long and short sides. The four corner points of the parking space frame are determined by the intersection of the dividing line and the parking space boundary line, rather than directly relying on the integrity of the point cloud. Even if some parking space line point clouds are incomplete due to occlusion by vehicles and pedestrians, a complete rectangular parking space frame can still be generated through geometric constraints, which enhances the robustness of the method in real complex scenarios.

[0020] (4) This invention eliminates the need for manual annotation. Through a complete technical process of point cloud processing and geometric fitting, it achieves fully automated processing from the input of the original point cloud to the output of the parking space-level map, solving the technical problems of traditional map generation relying on manual drawing, long annotation cycle, and delayed updates. Attached Figure Description

[0021] Figure 1 This is a flowchart of the automatic parking map generation method based on point cloud and semantic features according to the present invention; Figure 2 For the test point cloud, among which, Figure 2 (a) in the image represents the scan result. Figure 2 (b) in the image represents ground point clouds; Figure 3 This is a schematic diagram of point cloud voxelization. Figure 3 (a) Point cloud distribution before voxelization Figure 3 (b) Voxelized point cloud distribution; Figure 4 The straight line fitting is used for the continuous boundary line of the parking space. Figure 4 (a) in the figure is a schematic diagram of parallel constraint fitting of the present invention. Figure 4 (b) in the figure is a schematic diagram of the redundant linear filtering of the present invention; Figure 5Point cloud for the remaining dividing lines; Figure 6 Fit a straight line to the parking spot cloud; Figure 7 Visualize the parking space frame; Figure 8 This is the final editable parking map. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] Example 1 This embodiment uses a school parking lot as the experimental subject. The parking lot has 85 standard parking spaces and there are interference objects such as trees, street lamp poles, and buildings around the parking lot.

[0025] This embodiment uses the Pegasus Slam100 handheld 3D laser scanner to acquire raw point cloud data. This scanner is equipped with a 16-line LiDAR, a laser field of view of 270°×360°, a maximum ranging distance of 120m, a relative accuracy of 2cm, an absolute accuracy of 5cm, a point frequency of 320kpts / s, and a camera resolution of 5 megapixels. A total of 46,004,628 raw point clouds were acquired, containing complete scene information including the ground, vehicles, trees, and walls. The scan results are shown below. Figure 2 As shown in (a) of the diagram.

[0026] like Figure 1 As shown, the specific steps are as follows: Step S1: Point cloud preprocessing.

[0027] 1.1 SOR noise reduction.

[0028] The original point cloud contains outlier noise points caused by atmospheric particle interference, multipath echoes caused by diffuse reflection from object surfaces, weather effects, and instrument measurement errors. These outlier noise points significantly affect the accuracy of subsequent parking space line point cloud extraction. This embodiment uses Statistical Outlier Removal (SOR) filtering for denoising.

[0029] First, a KD-tree is constructed to accelerate the search for nearest neighbors. For each point in the point cloud... Calculate its to Neighboring points distance : (1); in, and Points and points coordinates This indicates the index of the point currently being calculated. Indicates the first point Nearest neighbor index. Calculation points To its The average distance of the nearest points : (2); The average distance of all points for: (3); Standard deviation of the mean distance of all points for: (4); in, This represents the total number of points in the point cloud.

[0030] This embodiment determines the SOR filtering parameters based on the actual point cloud density: when the number of neighboring points is less than 15, isolated noise is not completely removed; when it is greater than 25, it is easy to mistakenly delete parking space line edge points. Therefore, the number of neighboring points... Set to 20. Standard deviation coefficient. At this time, both the noise removal rate and the parking space line point retention rate reach their maximum. Therefore, this embodiment sets... , .

[0031] 1.2 Slope filtering separates ground points.

[0032] Even after denoising, the point cloud still contains non-ground points such as trees, vehicles, and buildings, requiring the separation of ground points to reduce the complexity of subsequent processing. This embodiment employs a slope-based ground filtering algorithm, which determines whether a point belongs to the ground by the slope difference (the ratio of height difference to distance) between the control point and its neighboring points.

[0033] The maximum permissible slope function is used as the filtering criterion. This function defines the maximum allowable height difference between two points. With horizontal distance Changes: (5); in, The maximum allowable slope angle is set. The horizontal distance between the two points To account for height offset tolerance, in this embodiment, based on the scanner accuracy and the height of ground undulations in the scene (the parking lot is a flat surface with undulations not exceeding 5cm), the height offset is adjusted accordingly. Set it to 0.02m.

[0034] Testing sites Find candidate ground points from its surrounding neighborhood. Set the judgment conditions: (6); in, , The heights of the two points are... Let be the Euclidean distance between the two points on the horizontal plane.

[0035] In this embodiment, since the slope of the parking lot surface area is less than 5°, while the slope of the obstacles would significantly exceed this range, a slope threshold is set. Search for its 10 nearest neighbors. Based on formulas (5) and (6), separate ground points from non-ground points, and retain ground points for subsequent processing. The results are as follows: Figure 2 As shown in (b) of the diagram.

[0036] Step S2: Rapid extraction of parking space line point cloud.

[0037] 2.1 Voxelization downsampling.

[0038] Point clouds acquired by lidar exhibit significant inhomogeneity, with some areas showing excessively dense or sparse point clouds. This embodiment employs a voxelization downsampling method to simplify the point cloud, such as... Figure 3 As shown.

[0039] The effectiveness of point cloud voxelization is related to the choice of voxelization edge length. Setting the edge length too large will lead to the loss of local point cloud features, while setting it too small will result in poor simplification. To both simplify the point cloud and preserve its overall distribution and local features, the voxelization edge length is typically set to 1.5 to 3 times the average distance of the point cloud. In this example, referencing the average density of parking lot point clouds (0.03 m / point), the voxel size is set to 0.05 m (approximately 1.5 times the average density).

[0040] 2.2 Parking space line point cloud extraction based on reflection intensity.

[0041] Existing methods for extracting parking line point clouds are mostly based on the color contrast between the parking lines and the ground, using the RGB color information of the point cloud for segmentation. However, these methods are greatly affected by lighting conditions and have poor anti-interference capabilities. This embodiment uses reflection intensity thresholding to extract parking line point clouds. The reflection intensity is determined by the target material and is unaffected by ambient lighting.

[0042] First, the extracted ground point cloud is projected onto the xoy plane in a spatial rectangular coordinate system; then, the mean value of the reflection intensity is calculated. and standard deviation : (7); (8); in, The number of points in the point cloud. For the first The reflection intensity value at each point.

[0043] Finally, point cloud binarization segmentation is performed using a threshold to determine the reflection intensity. Points with reflection intensity above the threshold are classified as belonging to the parking line point cloud, while points with reflection intensity below the threshold belong to the ground point cloud. (9); in, The standard deviation multiple is used here. In this embodiment, since the parking lot is a standard parking lot with no complex reflection scenes or severely worn parking spaces, and the reflection intensity of the ground point cloud follows a normal distribution, the standard deviation multiple is used. Set to the default value of 1. This threshold maximizes the extraction of parking space line points while minimizing the introduction of ground noise points.

[0044] Step S3: Line fitting of parking space line point cloud with geometric constraints.

[0045] 3.1 DBSCAN coarse clustering.

[0046] Although the obtained parking space line point cloud has initially identified the target area, it still exists as a discrete and unordered set of three-dimensional coordinates. In this embodiment, the DBSCAN algorithm is used to cluster consecutive parking spaces into the same class, and subsequent processing is performed separately on consecutive parking space lines within each class.

[0047] In this embodiment, the neighborhood search radius is set to 0.2m and the minimum number of points is 10, based on the actual point cloud spacing. After segmentation, effective clusters are extracted, and different clusters are color-coded.

[0048] 3.2 Fitting of continuous parking space boundary lines.

[0049] For the point cloud data of each cluster, this embodiment uses the RANSAC algorithm to fit the parking line point cloud. Two points are randomly selected from the point cloud. , Perform a line fitting and solve using the general formula: (10); (11); (12); (13); After obtaining the equation of the line, iterate through all points in the point cloud. Calculate the perpendicular distance from each point to the fitted two-dimensional line. : (14); Let the distance threshold be ,like If the distance is within a certain threshold, then the point is determined to be an interior point. In this embodiment, the straight line fitted by RANSAC is the geometric center line of the parking space line, and the distance threshold is used. The optimal width is 1 / 2 of the physical width of the parking space line. In the experiment, the actual physical width of the parking space line in the parking lot was 15cm, therefore... Take 7.5cm. Through continuous iteration, obtain multiple fitted two-dimensional straight lines as candidate lines. Compare the number of interior points of each candidate line. The line with the most interior points is taken as the main line of each major category of point cloud and serves as the reference for the boundary line of continuous parking spaces.

[0050] Since a single cluster may contain multiple parallel continuous parking space boundary lines, the point cloud after the straight line fitting is completed is no longer used for subsequent straight line fitting. The RANSAC fitting process is repeated on the remaining point cloud (excluding points inside the main line) to obtain candidate continuous parking space boundary lines. To ensure that the fitting results conform to the parallelism characteristics of continuous parking space boundary lines, an angle constraint method is used to calculate the angle between the candidate line and the main line. : (15); in, The slope of the principal line. Let the slope of the candidate line be the angle between the two lines. If the value is less than a set threshold, the candidate line is retained, its slope is adjusted to be the same as the main line, and the intercept is recalculated based on the midpoint coordinates of the points within the candidate line. Figure 4 (a) In this embodiment, the angle threshold is set to 15° (in other cases, less than 45° is acceptable).

[0051] To avoid redundant lines in the fitting results, the average distance of the lines within the point cloud boundary is calculated. Based on the parking space line width, lines with distances less than a threshold are filtered out, such as... Figure 4 As shown in (b) of the figure. In this embodiment, the distance threshold is set to 1m. This setting is based on the following considerations: in order to ensure that only one straight line is generated for a continuous boundary line of a parking space, the threshold can be greater than 15cm based on the actual line width of the parking space line of 15cm; however, in order to prevent some noise interference, the distance threshold should be set as large as possible, and at the same time, it should not exceed the length of the long side of a single parking space frame (5m), so 1m is taken as the filtering threshold.

[0052] 3.3. Separator line fitting.

[0053] After fitting the boundary lines of continuous parking spaces, the remaining dividing lines are fitted, such as... Figure 5As shown. This embodiment uses DBSCAN density clustering for fine clustering, grouping each dividing line into a smaller cluster. The parameters for fine clustering are set as follows: neighborhood radius is set to 1.5m, minimum number of points threshold is set to 8, and minimum effective length of short straight lines is set to 1m.

[0054] Based on the geometric perpendicularity between the parking space separator lines and the boundary lines of consecutive parking spaces, the RANSAC algorithm is used to fit a straight line to each valid class. After fitting, valid separator lines are filtered by angle deviation, the angle between the fitted line and the target slope is calculated, and only lines that meet the conditions are retained. The angle deviation threshold is set to 5°.

[0055] Meanwhile, to further optimize the fitting accuracy of the dividing line and ensure that the parking space line frame formed later is a standard rectangle, the slope of the selected straight line is unified, the slope of the dividing line is forcibly corrected to the target slope, and the intercept is optimized. Based on the midpoint coordinates of the interior points of the short side, the intercept is recalculated.

[0056] Ultimately, retain two or more valid edges, such as Figure 6 As shown.

[0057] Step S4: Parking lot map is automatically generated.

[0058] After fitting the boundary lines and dividing lines of continuous parking spaces, this embodiment constructs parking space frames through geometric relationships and displays them visually.

[0059] The four corner points of the parking space frame are determined by the intersection of the dividing line and the boundary line of the continuous parking spaces. To prevent interference between the fitted lines of different clusters, the intersection points are calculated separately for each class in the coarse cluster. The intersection points are obtained by solving the equations of the boundary line and the dividing line of the continuous parking spaces. ), which is the corner point of the parking space frame.

[0060] The intersections of each pair of adjacent dividing lines with the two parking space boundary lines form a rectangle. Taking two adjacent dividing lines as a group, the four intersections of each group of dividing lines with the two parking space boundary lines form a sequence of the rectangle's vertices. , , , To correctly draw the polygon, the four corner points need to be sorted in the order of bottom left, top left, top right, and bottom right. Then, the four vertices are connected to form a closed rectangle, which serves as the geometric representation of the parking space frame.

[0061] Each parking space is numbered and labeled based on its center point coordinates. The final output is a visualized map containing the parking space numbers, such as... Figure 7 As shown.

[0062] Finally, to ensure adjustability later, the generated map is exported as a DXF format, allowing for individual editing of each parking space. This facilitates subsequent parking space adjustments and numbering. Figure 8 .

[0063] Experimental results and analysis.

[0064] In this embodiment, the parking space detection accuracy reached 100%, with all 85 parking spaces identified without any missed or false detections. Both the parking space corner points and the fitted parking space lines were within the actual point cloud coverage area, and no boundary offset issues were observed.

[0065] Regarding light interference, this embodiment collects and tests data at noon, dusk, and night. Since the reflection intensity is not affected by light, the results obtained in the three scenarios are the same, and the accuracy of parking space detection remains 100%.

[0066] Regarding occlusion interference, since the algorithm of this invention generates parking space frames by taking the intersection points of straight lines according to spatial relationships, even if some parking space line point clouds are incomplete, parking space frames can still be generated.

[0067] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0068] Therefore, this invention adopts the above-mentioned automatic parking map generation method and system based on point cloud and semantic features. By extracting parking space line point clouds through reflection intensity threshold segmentation, it overcomes the influence of illumination changes on parking space line extraction. Through a two-step clustering and geometric constraint straight line fitting strategy, it solves the technical problem that continuous parking space boundaries cannot be effectively distinguished. It achieves automatic generation of parking maps with 100% parking space detection accuracy, 0% false detection rate, and 0% false detection rate, and maintains stable detection performance under different illumination conditions.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatically generating parking lot maps based on point cloud and semantic features, characterized in that, Includes the following steps: Step S1: Obtain the original LiDAR point cloud of the parking lot and preprocess the original LiDAR point cloud to obtain the ground point cloud; Step S2: Based on the obtained ground point cloud, extract the parking space line point cloud according to the reflection intensity of each point, and perform clustering and segmentation on the parking space line point cloud to obtain multiple independent parking space line point cloud clusters. Step S3: Perform straight line fitting on each parking space line point cloud cluster to obtain the parking space boundary line and parking space dividing line; Step S4: Generate parking space frames based on the intersection of parking space boundary lines and parking space separator lines, and output the parking lot map.

2. The automatic parking map generation method based on point cloud and semantic features according to claim 1, characterized in that, In step S1, the preprocessing includes: A statistical outlier removal filtering algorithm is used to denoise the original lidar point cloud in order to remove outlier noise points; A slope-based ground filtering algorithm is used to separate ground points from non-ground points in the denoised point cloud to obtain the ground point cloud.

3. The automatic parking map generation method based on point cloud and semantic features according to claim 2, characterized in that, In slope-based ground filtering algorithms, the maximum allowable height difference between two points is... The calculation is as follows: ； in, The maximum allowable slope angle is set. The horizontal distance between the two points This is for height offset tolerance.

4. The automatic parking map generation method based on point cloud and semantic features according to claim 1, characterized in that, In step S2, extracting the parking space line point cloud based on the reflection intensity of each point specifically includes: Statistically calculate the mean reflection intensity at each point in the ground point cloud. and standard deviation ; The reflection intensity threshold is determined based on the mean and standard deviation. : ； in, It is a multiple of the standard deviation; The reflection intensity is greater than the reflection intensity threshold. The point cloud was extracted as parking space line point cloud.

5. The automatic parking map generation method based on point cloud and semantic features according to claim 1, characterized in that, In step S2, the density-based clustering algorithm DBSCAN is used to cluster and segment the parking space line point cloud, dividing the spatially continuous parking space line point cloud into the same parking space line point cloud cluster.

6. The automatic parking map generation method based on point cloud and semantic features according to claim 1, characterized in that, Step S3, specifically performing straight line fitting on each parking space line point cloud cluster, includes: Step S31: Use the random sampling consensus algorithm to perform straight line fitting on the current parking space line point cloud cluster to obtain the main straight line, which serves as the parking space boundary line. Step S32: Remove the inner points corresponding to the main line from the current parking space line point cloud cluster, and repeatedly perform the random sampling consensus algorithm to fit candidate lines on the remaining point cloud; Step S33: Calculate the angle between the candidate line and the main line. If the angle is less than the angle threshold, the candidate line is retained as another parking space boundary line and its slope is adjusted to be the same as the main line. If the angle is greater than or equal to the angle threshold, the candidate line is discarded.

7. The automatic parking map generation method based on point cloud and semantic features according to claim 6, characterized in that, Step S3, which involves performing straight-line fitting on each parking space line point cloud cluster, also includes: After fitting all parking space boundary lines, density-based clustering algorithm is used again to finely cluster the dividing line point cloud from the remaining point cloud of the current parking space line point cloud cluster, resulting in multiple dividing line point cloud sub-clusters. For each sub-cluster of the dividing line point cloud, a straight line is fitted using a random sampling consensus algorithm to obtain the parking space dividing line, and the slope of the parking space dividing line is forced to satisfy the vertical geometric constraint with the slope of the parking space boundary line.

8. The automatic parking map generation method based on point cloud and semantic features according to claim 7, characterized in that, Step S4, generating the parking space frame specifically includes: Calculate the four intersection points of the two adjacent parking space separator lines and the two parking space boundary lines; Connect the four intersection points in spatial order to form a closed rectangle, which serves as the parking space frame.

9. The automatic parking map generation method based on point cloud and semantic features according to claim 1, characterized in that, In step S4, the output parking lot map contains a unique number for each parking space frame.

10. A parking lot map automatic generation system based on point cloud and semantic features, characterized in that, A method for automatically generating parking lot maps based on point cloud and semantic features as described in any one of claims 1-9, comprising: The point cloud acquisition and preprocessing module is used to acquire the original LiDAR point cloud of the parking lot and preprocess the original LiDAR point cloud to obtain the ground point cloud. The parking line extraction and clustering module is used to extract parking line point clouds based on ground point clouds and to cluster and segment the parking line point clouds to obtain multiple independent parking line point cloud clusters. The line fitting module is used to perform line fitting on each parking space line point cloud cluster to obtain the parking space boundary line and parking space dividing line. The map generation module is used to generate parking space frames based on the intersection of parking space boundary lines and parking space separator lines, and output a parking lot map.

Images