A road point cloud segmentation optimization method based on road centerline

Abstract

The present invention belongs to the field of computer vision and three-dimensional point cloud processing, and discloses a road point cloud segmentation optimization method based on the road centerline, including the following steps: model training and tuning: by optimizing the training process and model structure, the model's perception of road geometric characteristics is improved; road segmentation iterative optimization: using the optimized model to generate preliminary segmentation results, calculating the road centerline information based on the results, and guiding the model to segment again as prior knowledge, thereby improving the optimization segmentation accuracy and global connectivity; programmatic generation: based on the optimization results, a high-precision road model for road reconstruction is generated. The present invention improves the fine segmentation effect of road point clouds by introducing relative normal vectors and curvature encoding; by extracting the road centerline and width information and feeding it back to the segmentation model, the global connectivity of road segmentation is successfully optimized, ensuring the continuity and integrity of the segmentation results.

Description

Technical Field

[0001] The invention belongs to the field of computer vision and three-dimensional point cloud processing, and specifically relates to a road point cloud segmentation optimization method based on road centerline extraction. Background Art

[0002] With the in-depth advancement of smart city construction and the continuous development of 3D modeling technology, urban point cloud data, as an important data source for refined scene modeling and analysis, has gradually shown broad application prospects. Whether in urban planning, traffic navigation, or unmanned driving, high-precision point cloud processing technology has become the key. However, due to its own irregularity, sparsity and massive characteristics, urban point cloud data faces huge challenges in processing efficiency and analysis accuracy. Especially in road scenes, due to the complex surface characteristics of the road, fuzzy boundaries, and uneven distribution of point clouds, traditional processing methods are difficult to fully capture the geometric characteristics and topological relationships of the road. Although the prior geometric information such as centerline distribution and width contained in road point cloud data provides a potential theoretical basis for improving semantic segmentation accuracy and global connectivity, the current technical development is still insufficient in effectively utilizing this information.

[0003] In the prior art, mainstream methods mostly use deep learning networks such as PointNet, RandLA-Net and other models. Specifically, they have the following main shortcomings: 1) They rely too much on the XYZ coordinates and color information of the point cloud, and fail to make full use of the geometric structure features in the point cloud, resulting in poor performance in capturing local geometric details and modeling global topological relationships; 2) Most segmentation network optimization goals focus on classification accuracy, ignoring the importance of topological connectivity in road segmentation, which easily leads to segmentation results. Discontinuity or discontinuity; 3) The prior art fails to effectively utilize prior geometric information such as the road centerline, and misses the opportunity to optimize the segmentation results through geometric constraints. Based on the above shortcomings, the present invention proposes an iterative optimization strategy combined with road centerline geometric constraints, aiming to improve the accuracy and global connectivity of road point cloud segmentation, while solving the problem of insufficient capture of local geometric features, and providing technical support for efficient processing of smart city point cloud data and road modeling. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a road point cloud segmentation optimization method based on the road centerline, which realizes the global connectivity optimization and local detail completion of the point cloud semantic segmentation results, and provides accurate and reliable data support for road modeling, navigation and multi-scenario applications in smart cities.

[0005] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0006] A road point cloud segmentation optimization method based on road centerline, the specific process includes the following steps:

[0007] S1. Model training and tuning: Improve the model's perception and adaptability to road geometry by optimizing the training process and model structure;

[0008] S2, iterative optimization of road segmentation: using the model optimized in step S1 to generate a preliminary segmentation result, and calculating the road centerline information based on the result, which is used as prior knowledge to guide the model to segment again, thereby improving the optimization segmentation accuracy and global connectivity;

[0009] S3. Programmatic generation: Based on the optimization results in step S2, a high-precision road model for road reconstruction is generated.

[0010] Furthermore, the specific steps of step S1 are as follows:

[0011] S11. Multi-feature extraction: Calculate the normal vector and curvature information of each point from the point cloud data, and encode the relative changes of this information into features, cascade with the point cloud geometric features of the original network, learn point cloud features and segment point cloud data based on the new features after cascading, optimize the feature extraction module of the deep learning network, and enhance the accuracy and robustness of the point cloud segmentation model by learning more features;

[0012] S12, point cloud preprocessing: using a grid sampling method to perform preliminary downsampling on the point cloud data, and further refining the samples through farthest point sampling; after determining the centroid of the voxel, the grid sampling method selects the original three-dimensional point closest to the centroid to replace all the original three-dimensional points in the voxel, thereby reducing the impact of sampling offset;

[0013] S13. Hyperparameter tuning: Construct an optimization objective function and use the optimization algorithm to find the hyperparameters that optimize the model segmentation accuracy and connectivity performance.

[0014] The specific method of step S13 is as follows:

[0015] Through the Bayesian optimization method, a method with road connectivity as the core is constructed, and the weighted sum of road connectivity and segmentation correctness is used as the objective function. The average path length similarity (APLS) is calculated by combining graph theory metrics and Dijkstra's shortest path algorithm. The specific expression of the average path length similarity is as follows:

[0016]

[0017] Where N represents the number of paths, L(a,b) represents the length of the path (a,b) in the real graph, and L( a',b') represents the path in the prediction graph ( a',b' )length, a Point and b The points represent the starting node and the ending node selected from all the road nodes in the real road graph. a' and b' The points represent the predicted road map and a Point and b The node corresponding to the point.

[0018] Furthermore, the specific steps of step S2 are as follows:

[0019] S21, initial segmentation: using the multi-feature information extracted in step S1 to train and optimize the model, perform semantic segmentation on the input point cloud data, identify and extract the road point cloud to generate a preliminary road point cloud segmentation result;

[0020] S22, road centerline extraction and road width calculation: projecting the preliminary road point cloud segmentation result onto a two-dimensional plane to generate a road binary map, and then extracting the road centerline from the binary map to complete the calculation of the road width;

[0021] The specific method of step S22 is as follows:

[0022] S221, road refinement map: after the point cloud segmentation and projection complete the two-dimensional plane mapping, the morphological closing operation is used to close the holes caused by the object occlusion, and the Zhang-Suen algorithm is used for refinement to obtain the road refinement map;

[0023] S222, sub-road simplification processing: for each pixel point in the road refinement map, a neighborhood window of size 3×3 is used, and the pixels that are continuous along the counterclockwise direction of the window edge are a connected domain. If there are more than 2 connected domains, the pixel point is recorded as an intersection. After the intersection is located, the intersection is used as the demarcation point and the non-branched road section is the sub-road; then the neighborhood pixels of each intersection are deleted, and the pixels of each sub-road are divided based on the region growth to obtain the sub-road set, and the road centerline is processed by sub-road merging, sub-road extension and intersection point merging. Sub-road merging is used to merge multiple sub-roads with similar endpoints and similar directions. After the sub-road merging is completed, the sub-road extension is performed to extend each interrupted sub-road within the size range of the intersection until it intersects with other sub-roads. The intersection point merging is used to simplify the adjacent intersection points into a new intersection point, and the new intersection position is the average of all the original intersection positions;

[0024] The RDP (Ramer-Douglas-Peuker) algorithm is used to simplify subpaths;

[0025] S223, road centerline generation: connect the road intersections and sub-roads generated in step S222 one by one to generate a corresponding road centerline;

[0026] S224, road width calculation: according to the road centerline and road surface binary map obtained in step S223, the road width is calculated by using the adaptive limiting filtering method; for each sub-road, multiple sampling is performed to obtain road width data, and two adjacent sampling values ​​are compared. If the difference exceeds a preset threshold, it is considered to be interference and the sampling value is not retained; if it does not exceed the preset threshold, the sampling value is retained, and finally the average of the multiple retained sampling values ​​is calculated as the road width.

[0027] S23, guided re-segmentation optimization: After extracting the road centerline information and road width information, it is used as prior knowledge and fed back to the segmentation network to guide the network to re-segment and assign road labels, thereby improving the accuracy and connectivity of the road point cloud segmentation results.

[0028] The specific method of guiding the re-segmentation is as follows:

[0029] S231, determine the range: according to step S22, obtain the road centerline and road width information, map the information back to three dimensions, then map the centerline point coordinates to an integer grid, and project them to a two-dimensional plane according to step S22 to reversely calculate the positions in the three-dimensional space corresponding to all points in each grid, and generate point cloud data corresponding to the centerline;

[0030] S232, re-segmentation optimization: According to the geometric distance between each point in the point cloud data and the road centerline, a dynamic confidence construction method of the distance between the point and the centerline is adopted. The confidence factor belongs to the probability enhancement coefficient of the road category, and its expression is as follows:

[0031]

[0032] Where d represents the Euclidean distance between the point and the center line of the road, k represents the maximum proportion of confidence enhancement, p represents the smoothness and rate of control attenuation, and max_d represents the maximum distance of attenuation;

[0033] The road classification probability of each point in the point cloud data is adjusted according to the confidence factor. The expression is as follows:

[0034]

[0035] Among them, P(road) represents the road category classification probability of a certain point before adjustment, and P(road)' represents the road category classification probability after adjustment; after adjustment, the predicted probabilities of all categories are renormalized to ensure that the sum is 1. After the probability adjustment is completed, the point cloud classification results are re-divided according to the updated probability distribution to generate optimized segmentation results.

[0036] Furthermore, the specific steps of step S3 are as follows:

[0037] S31, data output: according to the re-segmentation result of step S23, use step S22 to complete the road centerline extraction and width information calculation, and save the segmented scene in the form of a .json format file in the order of each sub-road, the intersection point of the sub-road, and the width of the sub-road, to provide data support for subsequent programmatic generation;

[0038] S32, road generation: based on the road data generated in step S31, use Houdini software to write a programmed generation algorithm to regenerate a road model that conforms to the actual data based on the data.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The road segmentation optimization method based on point cloud data of the present invention effectively enhances the model's perception of complex road geometric features by introducing normal vectors and curvature encoding, thereby improving the fine segmentation effect of road point clouds; by extracting road centerline and width information and feeding it back to the segmentation model, the global connectivity of road segmentation is successfully optimized, the connection problems caused by noise and road discontinuities are solved, and the continuity and integrity of the segmentation results are ensured. This method can generate a more refined road network model in a complex urban environment, significantly reduce geometric errors, and significantly improve reconstruction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of point cloud data road segmentation of the present invention;

[0042] Figure 2 It is a schematic diagram for qualitative comparison of the point cloud data road segmentation results of the present invention. DETAILED DESCRIPTION

[0043] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0044] like Figure 1 As shown, a road point cloud segmentation optimization method based on the road centerline, the specific process includes the following steps:

[0045] S1. Model training and tuning: By optimizing the training process and model structure, the model's perception and adaptability to road geometry characteristics are improved. The specific steps are as follows:

[0046] S11. Multi-feature extraction: Calculate the normal vector and curvature information of each point from the point cloud data, and encode the relative changes of this information into features, cascade with the point cloud geometric features of the original network, learn point cloud features and segment point cloud data based on the new features after cascading, optimize the feature extraction module of the deep learning network, and enhance the accuracy and robustness of the point cloud segmentation model by learning more features;

[0047] S12, point cloud preprocessing: using a grid sampling method to perform preliminary downsampling on the point cloud data, and further refining the samples through farthest point sampling; after determining the centroid of the voxel, the grid sampling method selects the original three-dimensional point closest to the centroid to replace all the original three-dimensional points in the voxel, thereby reducing the impact of sampling offset;

[0048] S13. Hyperparameter tuning: Construct an optimization objective function and use the optimization algorithm to find the hyperparameters that optimize the model segmentation accuracy and connectivity performance.

[0049] The specific method of step S13 is as follows:

[0050] Through the Bayesian optimization method, we constructed a method with road connectivity as the core, using the weighted sum of road connectivity and segmentation correctness as the objective function, combining graph theory metrics with the Dijkstra shortest path algorithm to calculate the average path length similarity (APLS). The indicator is used to evaluate the similarity of graphs and quantify road connectivity; the APLS indicator quantifies the similarity between the basic fact graph G and the proposed graph G' by summing the optimal path length differences between nodes. Specifically, the APLS metric calculates the difference in path lengths between all possible source (A) and target (B) nodes, and for missing paths, the maximum proportional difference (1.0) is used for penalty to ensure that path missing will have a significant impact on the score.

[0051] The specific expression of the average path length similarity is as follows:

[0052]

[0053] Where C∈[0,1], N represents the number of paths, L(a,b) represents the length of the path (a,b) in the real graph, and L( a',b' ) represents the path in the prediction graph ( a',b' )length, a Point and b The points represent the starting node and the ending node selected from all the road nodes in the real road graph. a' and b' The points represent the predicted road map and a Point and b The node corresponding to the point, node a' represents the proposed graph whose node in the graph is closest to the position of the ground truth node a. If the path (a, b) does not exist, the maximum contribution value 1.0 is used.

[0054] To further reflect the actual importance of the transportation network, we introduced a weight adjustment mechanism based on betweenness centrality to the APLS metric. Betweenness centrality describes the frequency of a node as an intermediate node in all shortest paths. In this paper, high-centrality nodes (such as intersections and main roads) are more critical in connectivity, so paths missing such nodes will be punished more severely. That is, the path loss weight of high-betweenness centrality nodes is larger to emphasize the importance of transportation hubs; the path loss weight of low-betweenness centrality nodes is smaller to reduce excessive attention to edge nodes. Therefore, the APLS metric is more in line with the actual needs of road network segmentation and ensures the importance of correct road connection.

[0055] S2, iterative optimization of road segmentation: Use the model optimized in step S1 to generate preliminary segmentation results, and calculate the road centerline information based on the results, which is used as prior knowledge to guide the model to segment again, thereby improving the optimization segmentation accuracy and global connectivity. The specific steps are as follows:

[0056] S21, initial segmentation: using the multi-feature information (normal vector and curvature) extracted in step S1 to train and optimize the model, perform semantic segmentation on the input point cloud data, identify and extract the road point cloud to generate a preliminary road point cloud segmentation result;

[0057] S22, road centerline extraction and road width calculation: projecting the preliminary road point cloud segmentation result onto a two-dimensional plane to generate a road binary map, and then extracting the road centerline from the binary map to complete the calculation of the road width;

[0058] The specific method of step S22 is as follows:

[0059] S221, road refinement map: after the point cloud segmentation and projection complete the two-dimensional plane mapping, the morphological closing operation is used to close the holes caused by the object occlusion, and the Zhang-Suen algorithm is used for refinement to obtain the road refinement map;

[0060] S222, sub-road simplification processing: for each pixel point in the road refinement map, a neighborhood window of size 3×3 is used, and the pixels that are continuous in the counterclockwise direction along the edge of the window are a connected domain. If there are more than 2 connected domains, the pixel point is recorded as an intersection. After the intersection is located, the intersection is used as the demarcation point and the non-branched road section is used as a sub-road; then the neighborhood pixels of each intersection are deleted, and the pixels of each sub-road are divided based on region growth to obtain a sub-road set;

[0061] In order to solve the problem of straight roads being bent due to the obstruction of roadside trees, it is necessary to smooth the centerline of the road, that is, to replace the original pixel set sum with as few pixel sets as possible, so as to reduce the amount of stored data while smoothing the broken line. In order to solve the problem of road centerline deformation near intersections, the present invention processes the road centerline through sub-road merging, sub-road extension and intersection point merging. Sub-road merging is used to merge multiple sub-roads with similar endpoints and directions. After the sub-road merging is completed, the sub-road extension is performed to extend each interrupted sub-road within the size of the intersection until it intersects with other sub-roads. Intersection point merging is used to simplify close intersection points into a new intersection point, and the new intersection position is the average of all original intersection positions;

[0062] The RDP (Ramer-Douglas-Peuker) algorithm is used to simplify subpaths;

[0063] S223, road centerline generation: connecting the road intersections and sub-roads generated in step S222 one by one to generate a road centerline that meets expectations;

[0064] S224, road width calculation: According to the binary image of the road centerline and road surface obtained in step S223, due to the obstruction of trees, the scanned road width data often has noise. The present invention adopts an adaptive limiting filtering method to calculate the road width. The limiting filtering eliminates signals with large fluctuations, thereby eliminating the interference of random noise. For each sub-road, multiple samplings are performed to obtain road width data, and two adjacent sampling values ​​are compared. If the difference exceeds a preset threshold, it is determined to be interference and the sampling value is not retained; if it does not exceed the preset threshold, the sampling value is retained, and finally the average is calculated based on the multiple retained sampling values ​​as the road width.

[0065] S23, guided re-segmentation optimization: After extracting the road centerline information and road width information, it is used as prior knowledge and fed back to the segmentation network to guide the network to re-segment and assign road labels, thereby improving the accuracy and connectivity of the road point cloud segmentation results.

[0066] The specific method of guiding the re-segmentation is as follows:

[0067] S231, determine the range: according to step S22, obtain the road centerline and road width information, and when segmenting, the probability of data points within the road centerline and its width being classified as roads should be increased, and the information is mapped back to three dimensions, and then the centerline point coordinates are mapped to integer grids, and projected to a two-dimensional plane according to step S22 to reversely calculate the positions of all points in each grid in the three-dimensional space, and generate point cloud data corresponding to the centerline;

[0068] S232, re-segmentation optimization: According to the geometric distance between each point in the point cloud data and the road centerline, a dynamic confidence construction method of the distance between the point and the centerline is adopted. The confidence factor belongs to the probability enhancement coefficient of the road category, and its expression is as follows:

[0069]

[0070] Wherein, d represents the Euclidean distance between the point and the center line of the road, k represents the maximum proportion of confidence enhancement, p represents the smoothness and rate of control attenuation, and max_d represents the maximum distance of attenuation, that is, the width of the road obtained in step S22;

[0071] The road classification probability of each point in the point cloud data is adjusted according to the confidence factor. The expression is as follows:

[0072]

[0073] Among them, P(road) represents the probability of road classification of a certain point before adjustment, and P(road)' represents the probability of road classification after adjustment; after adjustment, the predicted probabilities of all categories are renormalized to ensure that the sum is 1. After the probability adjustment is completed, the point cloud classification results are re-divided according to the updated probability distribution to generate optimized segmentation results. This adjustment method based on dynamic confidence can effectively improve the accuracy of road category segmentation, especially for the processing of boundary areas and low-confidence points.

[0074] S3, programmatic generation: Based on the optimization results in step S2, a high-precision road model for road reconstruction is generated. The specific steps are as follows:

[0075] S31, data output: according to the re-segmentation result of step S23, use step S22 to complete the road centerline extraction and width information calculation, and save the segmented scene in the form of a .json format file in the order of each sub-road, the intersection point of the sub-road, and the width of the sub-road, to provide data support for subsequent programmatic generation;

[0076] S32, road generation: based on the road data generated in step S31, use Houdini software to write a programmed generation algorithm to regenerate a road model that conforms to the actual data based on the data. Example

[0077] First, the SensatUrban dataset is selected for the case study. This dataset covers an urban area of ​​6 square kilometers, including parts of Cambridge and Birmingham, with a total of 2.847 billion points; the SensatUrban dataset contains many challenging roads, including roads of different widths and types, roads blocked by trees or cars, and various complex intersection shapes.

[0078] A road point cloud segmentation optimization method based on road centerline, the specific process includes the following steps:

[0079] S1. Model training and tuning: By optimizing the training process and model structure, the model's perception and adaptability to road geometry features are improved to obtain better segmentation results.

[0080] S2, iterative optimization of road segmentation: using the model optimized in step S1 to generate a preliminary segmentation result, and calculating the road centerline information based on the result, which is used as prior knowledge to guide the model to segment again, thereby improving the optimization segmentation accuracy and global connectivity;

[0081] To evaluate the accuracy of the semantic segmentation results of point clouds, the overall accuracy (OA), mean intersection over union (mIoU), and mean accuracy (mAcc) are generally used as evaluation criteria; the overall accuracy (OA) indicates the proportion of correctly classified points to the total number of points, the intersection over union (IoU) indicates the ratio of the intersection and union of the true area and the predicted area, and the mean IoU calculates the average ratio of the intersection and union of the predicted area and the true area of ​​each category. The average IoU provides a more balanced performance metric that takes into account the classification performance of each category and will not be biased by the large number of pixels in a certain category. The average accuracy indicates the average ratio of the number of correctly classified points in each category to all points in that category. mAcc gives equal importance to each category, which helps to evaluate the performance of the model on small categories, but it may ignore the imbalance problem between categories.

[0082] In order to make a fair comparison with other networks on the SensatUrban dataset, the training parameters of each comparison algorithm are consistent, the training cycle (epoch) is set to 50, the batch size (batch size) is set to 6, and the SensatUrban dataset is divided into 31 blocks for training, 6 blocks for validation, and 6 blocks for test. Table 1 shows the performance of different point cloud segmentation methods on the SensatUrban dataset under three evaluation indicators.

[0083] Table 1 Schematic diagram of quantitative comparison of point cloud data road segmentation results of the present invention

[0084]

[0085] As shown in Table 1, in terms of the accuracy of road point cloud segmentation, the model of the present invention exceeds the existing mainstream methods in multiple indicators, specifically, the overall accuracy (OA) reaches 95.2%, the average classification accuracy (mAcc) is 90.5%, and the average intersection over union (mIoU) is 83.5%, which is significantly improved compared with the traditional method;

[0086] S3. Programmatic generation: Based on the optimization results, a high-precision road model for road reconstruction is generated.

[0087] like Figure 2 As shown, for the original point cloud data, after the road centerline and width information are generated by step S2, the high-quality road reconstruction rendering generated by the program in step S3 clearly reflects the accurate geometric structure and details of the road, proving its strong advantages in practical applications.

[0088] The above description is only an implementation mode of the present invention. It is stated again that for ordinary technicians in this technical field, several improvements can be made to the present invention without departing from the principle of the present invention. These improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A road point cloud segmentation optimization method based on road centerline, characterized by: The specific process includes the following steps: S1. Model training and tuning: Improve the model's perception and adaptability to road geometry by optimizing the training process and model structure; S2, iterative optimization of road segmentation: using the model optimized in step S1 to generate a preliminary segmentation result, and calculating the road centerline information based on the result, which is used as prior knowledge to guide the model to segment again, thereby improving the optimization segmentation accuracy and global connectivity; S3, programmatic generation: based on the optimization result in step S2, a high-precision road model for road reconstruction is generated; The specific steps of step S1 are as follows: S11. Multi-feature extraction: Calculate the normal vector and curvature information of each point from the point cloud data, and encode the relative changes of this information into features, cascade with the point cloud geometric features of the original network, learn the point cloud features and segment the point cloud data based on the new features after cascading, and optimize the feature extraction module of the deep learning network; S12, point cloud preprocessing: using a grid sampling method to perform preliminary downsampling on the point cloud data, and further refining the samples by farthest point sampling; after determining the centroid of a voxel, the grid sampling method selects the original three-dimensional point closest to the centroid to replace all the original three-dimensional points in the voxel; S13, Hyperparameter tuning: Construct an optimization objective function and use the optimization algorithm to find the hyperparameters that achieve the best segmentation accuracy and connectivity performance of the model; The specific steps of step S2 are as follows: S21, initial segmentation: using the multi-feature information extracted in step S1 to train and optimize the model, perform semantic segmentation on the input point cloud data, identify and extract the road point cloud to generate a preliminary road point cloud segmentation result; S22, road centerline extraction and road width calculation: projecting the preliminary road point cloud segmentation result onto a two-dimensional plane to generate a road binary map, and then extracting the road centerline from the binary map to complete the calculation of the road width; S23, guided re-segmentation optimization: After extracting the road centerline information and road width information, it is used as prior knowledge and fed back to the segmentation network to guide the network to re-segment and assign road labels, thereby improving the accuracy and connectivity of the road point cloud segmentation results.

2. The road point cloud segmentation optimization method based on road centerline according to claim 1, characterized in that: The specific steps of step S3 are as follows: S31, data output: according to the re-segmentation result of step S23, use step S22 to complete the road centerline extraction and width information calculation, and save the segmented scene in the form of a .json format file in the order of each sub-road, the intersection point of the sub-road, and the width of the sub-road, to provide data support for subsequent programmatic generation; S32, road generation: based on the road data generated in step S31, use Houdini software to write a programmed generation algorithm to regenerate a road model that conforms to the actual data based on the data.

3. The road point cloud segmentation optimization method based on road centerline according to claim 1, characterized in that: The specific method of step S13 is as follows: Through the Bayesian optimization method, a method with road connectivity as the core is constructed, and the weighted sum of road connectivity and segmentation correctness is used as the objective function. The average path length similarity is calculated by combining graph theory metrics and Dijkstra's shortest path algorithm. The specific expression of the average path length similarity is as follows: Where N is the number of paths and L(a,b) is the length of the path (a,b).

4. The road point cloud segmentation optimization method based on road centerline according to claim 1, characterized in that: The specific method of step S22 is as follows: S221, road refinement map: after the point cloud segmentation and projection complete the two-dimensional plane mapping, the morphological closing operation is used to close the holes caused by the object occlusion, and the Zhang-Suen algorithm is used for refinement to obtain the road refinement map; S222, sub-road simplification processing: for each pixel point in the road refinement map, a neighborhood window of size 3×3 is used, and the pixels that are continuous along the counterclockwise direction of the window edge are a connected domain. If there are more than 2 connected domains, the pixel point is recorded as an intersection. After the intersection is located, the intersection is used as the demarcation point and the non-branched road section is used as a sub-road; then the neighborhood pixels of each intersection are deleted, and the pixels of each sub-road are divided based on the region growth to obtain a sub-road set, and the road centerline is processed by sub-road merging, sub-road extension and intersection point merging; Sub-road merging is used to merge multiple sub-roads with similar endpoints and directions. After the sub-road merging is completed, the sub-road extension is performed to extend each interrupted sub-road within the size of the intersection until it intersects with other sub-roads. Intersection point merging is used to simplify the adjacent intersection points into a new intersection point, and the new intersection position is the average of all the original intersection positions. The RDP algorithm is used to simplify subpaths; S223, road centerline generation: connect the road intersections and sub-roads generated in step S222 one by one to generate a corresponding road centerline; S224, road width calculation: according to the road centerline and road surface binary image obtained in step S223, the road width is calculated by using the adaptive limiting filter method; for each sub-road, multiple sampling is performed to obtain road width data, and two adjacent sampling values ​​are compared. If the difference exceeds a preset threshold, it is determined to be interference and the sampling value is not retained; If it does not exceed the preset threshold, the sample value is retained, and finally the average is calculated based on the multiple retained sample values ​​as the road width.

5. The road point cloud segmentation optimization method based on road centerline according to claim 1, characterized in that: The specific method of guiding the re-segmentation is as follows: S231, determine the range: according to step S22, obtain the road centerline and road width information, map the information back to three dimensions, then map the centerline point coordinates to an integer grid, and project them to a two-dimensional plane according to step S22 to reversely calculate the positions in the three-dimensional space corresponding to all points in each grid, and generate point cloud data corresponding to the centerline; S232, re-segmentation optimization: According to the geometric distance between each point in the point cloud data and the road centerline point cloud, a dynamic confidence construction method of the distance between the point and the centerline is adopted. The confidence factor belongs to the probability enhancement coefficient of the road category, and its expression is as follows: Where d represents the Euclidean distance between the point and the center line of the road, k represents the maximum proportion of confidence enhancement, p represents the smoothness and rate of control attenuation, and max_d represents the maximum distance of attenuation; The road classification probability of each point in the point cloud data is adjusted according to the confidence factor. The expression is as follows: P(road)'=P(road)·(1+Confidence(d)) After the adjustment, the predicted probabilities of all categories are renormalized. After the probability adjustment is completed, the point cloud classification results are re-divided according to the updated probability distribution to generate optimized segmentation results.

Images