A road scene type recognition method and system based on vehicle-mounted laser point cloud
Patent Information
- Application Number
- CN202111452174.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-30
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-11-30
Smart Images

Figure CN113989784B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of road scene recognition, and in particular relates to a road scene type recognition method and system based on vehicle-mounted laser point cloud. Background Art
[0002] As a rapidly developing new surveying and mapping technology, the vehicle-mounted laser scanning system can quickly and accurately obtain spatial information of road scenes and surrounding objects, and is widely used in urban traffic management, smart transportation, high-precision maps and other fields. High-precision maps are the prerequisite for realizing autonomous driving. As an important part of static high-precision maps, the accurate identification of road boundary scene types helps to assist autonomous driving vehicles in path planning and decision-making.
[0003] At present, road scene recognition is mainly divided into two methods: image-based road scene type recognition method and GPS trajectory data-based road scene type recognition method. Image-based road scene type recognition method mainly extracts, describes and matches features from images, and recognizes intersection types by summarizing the semantic features, spatial structures and topological shapes of different intersection types. However, the intersection type recognition realized by this method relies on artificially designed features, and the recognition effect is seriously dependent on the accuracy of feature item design. In addition, the image-based method is sensitive to lighting and has difficulty in identifying complex road scene types. Based on GPS trajectory data, it mainly uses vehicle spatiotemporal trajectory data, constructs an intersection feature set based on the turning angle, speed, and dwell time at the intersection of the trajectory data, and classifies different intersection types based on different intersection type features. The GPS trajectory data used by this method is based on the trajectory area left by the vehicle. For roads with few vehicles, it is easy to cause incomplete trajectory data and it is difficult to accurately identify road scenes.
[0004] Patent CN108877267A discloses a method for detecting intersections based on a vehicle-mounted monocular camera, which designs a comprehensive loss function for intersection type classification and distance estimation loss functions, constructs a feature encoding subnetwork, an intersection type classification subnetwork, and a distance classification subnetwork to form an intersection type detection network, and uses a road color image captured by a vehicle-mounted monocular camera as input to the network, outputting the type of intersection on the road. However, this method is based on the monocular image acquired by the vehicle-mounted monocular camera, which has a low resolution and contains limited information, making it difficult to apply to large-scale road scenes.
[0005] Patent CN106896353A discloses a method for detecting intersections of unmanned vehicles based on three-dimensional laser radar, which uses the laser radar mounted on the unmanned vehicle to collect road scene point cloud data, converts the point cloud data into a raster image to extract multiple frames of height information maps, uses the height information map pixel points as feature vectors, and searches for intersection points in front of the unmanned vehicle based on the actual location of the unmanned vehicle and the correspondence between the real map, and finally inputs the height information map feature vector and the intersection point based on the support vector machine to classify the intersection type where the unmanned vehicle is located. However, this method converts the point cloud into a raster image method, and the generated image still has a loss of accuracy, and the classification accuracy depends on the selection of features, which affects the classification accuracy of road scene types.
[0006] Patent CN109271858A discloses a method for identifying intersections based on driving trajectory and visual lane edge data, which extracts a set of trajectory points with zero speed based on driving trajectory data, generates an intersection stop line based on the trajectory point set based on a density clustering algorithm, and extracts the intersection position based on the cut between the stop line data and the lane edge data of the camera associated with the road. However, this method can only identify the intersection area and fails to distinguish different intersection types.
[0007] Patent CN110688958A discloses a method for identifying intersections based on the GoogLeNet neural network, which uses road network data as intersection samples, inputs intersection samples into the GoogLeNet convolutional neural network training, and identifies different types of intersections. However, the intersections identified by this method can only identify scenes of urban trunk roads, and it is difficult to identify types such as high-speed forks and non-forks.
[0008] Patent CN109635722A discloses a high-resolution remote sensing image intersection automatic recognition method, which performs rough extraction of road and non-road categories based on remote sensing images, extracts road skeleton intersections as candidate intersection coordinates, and then identifies intersections. However, this method is only applicable to urban trunk roads with obvious road area characteristics. It is difficult to identify intersections on secondary roads, and the identified intersections fail to identify the intersection type.
[0009] Patent CN109815993A discloses a method for regional feature extraction, database establishment and intersection recognition based on GPS trajectory, which extracts mobile GPS trajectory based on the speed change of trajectory data, extracts road features using feature extraction window, and finally recognizes intersections based on KNN algorithm of sliding window. However, this method is difficult to identify areas with less trajectory data, and fails to distinguish the types of recognized intersections. Summary of the invention
[0010] The object of the present invention is to provide a method and system for identifying road scene types based on vehicle-mounted laser point clouds, which are conducive to identifying road boundaries of different road scene types.
[0011] To achieve the above object, the technical solution adopted by the present invention is: a road scene type recognition method based on vehicle-mounted laser point cloud data, comprising the following steps:
[0012] S1. Based on the vehicle-mounted laser point cloud data, obtain the ground point cloud and then obtain the road boundary point cloud;
[0013] S2, calculate the main direction of the road boundary point, take the driving direction of the vehicle-mounted laser scanning system as the forward direction of the trajectory point, calculate the current trajectory point vector, calculate the angle between the main direction and the trajectory point vector, calculate the characteristic value of the road boundary angle in the neighborhood of the trajectory point for each trajectory point, and classify the road straight segment and intersection segment according to the different angle values;
[0014] S3. Cluster the intersection segments into independent objects, and use the dynamic graph convolutional neural network DGCNN to classify intersections and T-junctions to obtain straight segments, intersections, and T-junctions.
[0015] Furthermore, the step S1 specifically includes the following steps:
[0016] S11. Based on the vehicle-mounted laser point cloud data, the ground point cloud is obtained by using the cloth simulation filtering method;
[0017] S12, based on the ground point cloud obtained in step S11, the ground point cloud is segmented using a supervoxel method, each point of the supervoxel has the same semantic label, and the road boundary point clouds with similar normals are segmented into the same semantic label, and all the road boundary point clouds in the label form a voxel block;
[0018] S13, calculate the average angle Δθ between the normal vector of the point cloud in each voxel block of the ground point cloud supervoxel and the Z axis of the point cloud data coordinate system, calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block, and calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block according to the voxel block normal threshold t normal and the voxel block elevation difference threshold t height , calculate to satisfy t normal ,t height The extracted voxel block is the voxel block containing the road boundary point cloud;
[0019] S14, based on the trajectory of the vehicle-mounted laser scanning system, retaining data points within a set range on both sides of the road, and removing data points outside the set range;
[0020] S15, cluster the roughly extracted road boundary voxel blocks using Euclidean clustering, and set the cluster point number threshold t num, remove the clusters with points less than the threshold number and obtain the denoised voxel blocks, that is, the road boundary point cloud.
[0021] Furthermore, the step S2 specifically includes the following steps:
[0022] S21, based on the acquired road boundary point cloud, using principal component analysis method to calculate the main direction of the road boundary point;
[0023] Input the road boundary point cloud and randomly select a point p in the point cloud i , retrieve p i Point cloud p in neighborhood r 1 、p 2 、p 3 …p n , calculate the mean of the x and y directions of the point cloud within the neighborhood r Calculate the covariance matrix C of the neighborhood points and select the eigenvector corresponding to the maximum eigenvalue in the covariance matrix as the road boundary point p i The main directions of the x and y directions Satisfies the following formula:
[0024]
[0025]
[0026]
[0027]
[0028] Among them, n is the number of point clouds in the point cloud neighborhood; Represents a point x in the neighborhood i ,y i The average value of the direction, x i ,y i Respectively represent the coordinate values of the i-th point cloud in the x direction and the y direction; Cov(x,y) represents the x, y covariance matrix;
[0029] S22, with the vehicle-mounted laser scanning system trajectory point t i As the center, establish a local search area ROI and calculate the initial trajectory point t within the ROI area i and the end trajectory point t j The direction vector v i , calculate v in the ROI area i and road boundary point p 1 、p 2 、p 3 …p n main direction The angle θ 1 ,θ 2 ,θ3 …θ n , this angle is called the angle characteristic value of the road boundary point, which satisfies the following formula:
[0030]
[0031]
[0032]
[0033] S23, using the vehicle-mounted laser scanning system trajectory point t i As the center, establish local ROI point by track point i , local ROI statistics i The eigenvalue of the angle at the inner boundary point θ i The cumulative sum Calculate point by point Traverse all trajectory points in turn and calculate the ROI corresponding to all trajectory points i The cumulative sum of the angle eigenvalues of the road boundary points within the region Intersection area ROI i Cumulative sum of angle eigenvalues Much larger than the sum of the eigenvalues of the angles in the straight line segment area Set the threshold θ threshold Separate straight line segments from intersection regions: Greater than θ threshold , then the current ROI i The area is a point cloud of intersections, otherwise it is a point cloud of straight line segments, which satisfies the following formula:
[0034]
[0035]
[0036] Furthermore, the step S3 specifically includes the following steps:
[0037] S31, based on the intersection point cloud extracted in S23, cluster the intersection point cloud into independent objects based on the connected branch clustering algorithm, where the cluster distance is C cluster , Based on the clustered independent objects at the intersection, the DGCNN network is used to classify the intersection type;
[0038] S32. The intersection object is sampled into n points using the farthest point and input into the DGCNN network. The data format is n×3, where 3 represents the initial number of features in the feature dimension of the point cloud. The input data is calculated through 3 layers of Edge-Conv and 1 layer of multi-layer perceptron MLP to obtain n×1024 high-dimensional features of the sampling points. The semantically similar points are aggregated to extract local feature information at different levels of the point cloud. Finally, the maximum pooling is used to obtain n×1 global features, and the extracted global features are assigned to each point. The features of different levels extracted by each EdgeConv are spliced. All feature information is fused through MLP and the high-dimensional features are reduced in dimensionality. Finally, 2×1 is output to represent the probability value corresponding to the current object being classified into 2 categories by the network, so as to determine the category of the intersection object and realize the classification of crossroads and T-junctions.
[0039] The present invention also provides a road scene type recognition system based on vehicle-mounted laser point cloud data, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above-mentioned method steps can be implemented.
[0040] Compared with the prior art, the present invention has the following beneficial effects: it breaks the limitation of using trajectory data or remote sensing data for the traditional classification of intersection types, directly faces the road boundary of the vehicle-mounted laser point cloud, identifies different road scene types, and identifies the road intersection types in different urban areas and high-speed scenes. The present invention first calculates the angle between the main direction of the road boundary and the direction of the trajectory line of the vehicle-mounted scanning system, classifies the straight line segments and intersection areas based on the similarities and differences between the angles at the intersection and the straight line area, clusters the independent objects in the intersection data area, and uses a deep neural network to classify the crossroads and T-junctions based on the clustered intersection areas, thereby realizing the recognition of the road boundary intersection type. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A flowchart of a method implementation of an embodiment of the present invention;
[0042] Figure 2 Schematic diagram of super-voxel segmentation of ground point cloud in an embodiment of the present invention;
[0043] Figure 3 This is a schematic diagram of a road boundary extraction result in an embodiment of the present invention;
[0044] Figure 4 Schematic diagram of the main direction of the road boundary in an embodiment of the present invention;
[0045] Figure 5 Schematic diagram of characteristic values of road boundary angles in an embodiment of the present invention;
[0046] Figure 6 Schematic diagram of the sum of angle characteristic values of the road boundary trajectory point area in an embodiment of the present invention;
[0047] Figure 7 Schematic diagram of the classification results of intersections and straight segments in an embodiment of the present invention;
[0048] Figure 8 This is a schematic diagram of the intersection type classification result in an embodiment of the present invention;
[0049] Fig. 9 This is a diagram of the DGCNN model architecture in an embodiment of the present invention. DETAILED DESCRIPTION
[0050] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0051] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.
[0052] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0053] like Figure 1 As shown, this embodiment provides a road scene type recognition method based on vehicle-mounted laser point cloud data, comprising the following steps:
[0054] like Figure 1 As shown, this embodiment provides a road scene type recognition method based on vehicle-mounted laser point cloud data, comprising the following steps:
[0055] S1. Based on the vehicle-mounted laser point cloud data, obtain the ground point cloud, and then obtain the road boundary point cloud.
[0056] S2. Calculate the main direction of the road boundary point, take the driving direction of the vehicle-mounted laser scanning system as the forward direction of the trajectory point, calculate the current trajectory point vector, calculate the angle between the main direction and the trajectory point vector, calculate the characteristic value of the road boundary angle in the neighborhood of the trajectory point for each trajectory point, and classify the road straight segment and intersection segment according to the different angle values.
[0057] S3. Cluster the intersection segments into independent objects, and use the dynamic graph convolutional neural network (DGCNN) to classify intersections and T-junctions to obtain straight segments, intersections, and T-junctions.
[0058] The specific implementation method of step S1 is:
[0059] S11. Based on the vehicle-mounted laser point cloud data, the cloth simulation filtering method is used to obtain the ground point cloud.
[0060] S12, based on the ground point cloud obtained in step S11, the ground point cloud is segmented using the supervoxel method. Each point of the supervoxel has the same semantic label. The road boundary point clouds with similar normals are segmented into the same semantic label. Here, similar normals refer to the normal angle between two road boundary point clouds being less than a set ε value. All road boundary point clouds in the label form a voxel block, which is represented as a voxel block in the label. Figure 2 shown.
[0061] S13, calculate the average angle Δθ between the normal vector of the point cloud in each voxel block of the ground point cloud supervoxel and the Z axis of the point cloud data coordinate system, calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block, and calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block according to the voxel block normal threshold t normal and the voxel block elevation difference threshold t height , calculate to satisfy t normal ,t height The extracted voxel block is the voxel block containing the road boundary point cloud.
[0062] S14. Based on the trajectory of the vehicle-mounted laser scanning system, retain the data points within the set range on both sides of the road, and remove the data points outside the set range.
[0063] S15, cluster the roughly extracted road boundary voxel blocks using Euclidean clustering, and set the cluster point number threshold t num , remove the clusters with less than the threshold number of points, and get the denoised voxel blocks, that is, the road boundary point cloud, such as Figure 3 shown.
[0064] The specific implementation method of step S2 is:
[0065] S21. Based on the acquired road boundary point cloud, a principal component analysis method is used to calculate the main direction of the road boundary points.
[0066] Input the road boundary point cloud and randomly select a point p in the point cloud i , retrieve p i Point cloud p in neighborhood r 1 、p 2 、p 3 …p n , calculate the mean of the x and y directions of the point cloud within the neighborhood r Calculate the covariance matrix C of the neighborhood points and select the eigenvector corresponding to the maximum eigenvalue in the covariance matrix as the road boundary point p iThe main directions of the x and y directions The main direction of the road boundary is Figure 4 As shown, the following formula is satisfied:
[0067]
[0068]
[0069]
[0070]
[0071] Among them, n is the number of point clouds in the point cloud neighborhood; Represents a point x in the neighborhood i ,y i The average value of the direction, x i ,y i Respectively represent the coordinate values of the i-th point cloud in the x direction and the y direction; Cov(x,y) represents the x, y covariance matrix.
[0072] S22, with the vehicle-mounted laser scanning system trajectory point t i As the center, establish a local search area ROI and calculate the initial trajectory point t within the ROI area i and the end trajectory point t j The direction vector v i , calculate v in the ROI area i and road boundary point p 1 、p 2 、p 3 …p n main direction The angle θ 1 ,θ 2 ,θ 3 …θ n , this angle is called the angle characteristic value of the road boundary point, such as Figure 5 As shown, the following formula is satisfied:
[0073]
[0074]
[0075]
[0076] S23, using the vehicle-mounted laser scanning system trajectory point t i As the center, establish local ROI point by track point i , local ROI statistics i The eigenvalue of the angle at the inner boundary point θ i The cumulative sum Calculate point by point Traverse all trajectory points in turn and calculate the ROI corresponding to all trajectory points i The cumulative sum of the angle eigenvalues of the road boundary points within the region like Figure 6 As shown, the intersection area ROI i Cumulative sum of angle eigenvalues Much larger than the sum of the eigenvalues of the angles in the straight line segment area Set the threshold θ threshold Separate straight line segments from intersection regions: Greater than θ threshold , then the current ROI i The area is a point cloud of intersections, otherwise it is a point cloud of straight line segments, such as Figure 7 As shown, the following formula is satisfied:
[0077]
[0078]
[0079] The specific implementation method of step S3 is:
[0080] S31, based on the intersection point cloud extracted in S23, cluster the intersection point cloud into independent objects based on the connected branch clustering algorithm, where the cluster distance is C cluster , based on the clustered independent objects at the intersection, the DGCNN network is used to classify the intersection type.
[0081] S32, the intersection object is sampled as n points using the farthest point, and the input is as follows Fig. 9 The DGCNN network shown in the figure has a data format of n×3, where 3 represents the initial number of features in the feature dimension of the point cloud. The input data is calculated through 3 layers of Edge-Conv and 1 layer of Multi-Layer Perception (MLP) to obtain n×1024 high-dimensional features of the sampling points, aggregate semantically similar points, extract local feature information at different levels of the point cloud, and finally use maximum pooling to obtain n×1 global features. The extracted global features are assigned to each point, and the features of different levels extracted by each EdgeConv are spliced. All feature information is fused through MLP and the high-dimensional features are reduced in dimension. Finally, 2×1 is output to represent the probability value corresponding to the current object being classified into 2 categories by the network, determine the category of the intersection object, and realize the classification of crossroads and T-junctions, such as Figure 8 shown.
[0082] This embodiment also provides a road scene type recognition system for implementing the above method, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above method steps can be implemented.
[0083] The present invention provides a road scene type recognition scheme based on vehicle-mounted laser point cloud data. First, based on the vehicle-mounted laser point cloud, the ground point cloud is extracted based on cloth simulation filtering, the road boundary points are clustered using the boundary-enhanced supervoxel method, the normal vectors in the voxel block are re-weighted, the road boundary points are extracted using the normal vector threshold and the voxel block elevation difference threshold, the discrete point cloud far away from the center of the trajectory line is eliminated, and the discrete point cloud less than the number of points is eliminated based on the connected branch method to obtain the road boundary point cloud. Based on the road boundary, the main direction is calculated, and the trajectory line of the vehicle-mounted laser scanning system is used as the forward direction, the angle value between the trajectory line driving direction and the main direction of the road boundary is calculated, the change of the angle value in the corresponding area of different trajectory points is counted, the angle values in different areas are sorted, and the top ten are selected as thresholds to classify the intersection area and the non-intersection area; based on the extracted intersection area, the connected branch clustering is used as an independent object, and the dynamic graph neural network classification network model is used to classify the intersection and T-intersection, and the T-intersection, intersection, and straight line segment are obtained by combining the classified straight line segments.
[0084] Compared with the prior art, the present invention directly constructs a road scene type recognition method for the road boundaries of vehicle-mounted laser point clouds, classifies intersection points and non-intersection points by calculating the changes in the direction angle values between the main direction of the road boundary and the trajectory line, clusters intersection data into independent objects, uses the DGCNN model to classify intersections and T-junctions, and finally outputs road boundaries of different road scene types, providing a new research method for the recognition of road scene types of vehicle-mounted laser point clouds.
[0085] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0086] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0087] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0088] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0089] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.
Claims
1. A road scene type recognition method based on vehicle-mounted laser point cloud data, It is characterized in that The following steps are involved: S1. Based on the vehicle-mounted laser point cloud data, obtain the ground point cloud and then obtain the road boundary point cloud; S2, calculate the main direction of the road boundary point, take the driving direction of the vehicle-mounted laser scanning system as the forward direction of the trajectory point, calculate the current trajectory point vector, calculate the angle between the main direction and the trajectory point vector, calculate the characteristic value of the road boundary angle in the neighborhood of the trajectory point for each trajectory point, and classify the road straight segment and intersection segment according to the different angle values; S3, clustering the intersection segments into independent objects, using the dynamic graph convolutional neural network DGCNN to classify the intersections and T-junctions, and obtaining straight segments, intersections and T-junctions; The step S2 specifically includes the following steps: S21, based on the acquired road boundary point cloud, using principal component analysis method to calculate the main direction of the road boundary point; Input the road boundary point cloud and randomly select a point p in the point cloud i , retrieve p i Point cloud p in neighborhood r 1 、p 2 、p 3 …p n , calculate the mean of the x and y directions of the point cloud within the neighborhood r Calculate the covariance matrix C of the neighborhood points and select the eigenvector corresponding to the maximum eigenvalue in the covariance matrix as the road boundary point p i The main directions of the x and y directions Satisfies the following formula: Among them, n is the number of point clouds in the point cloud neighborhood; Represents a point x in the neighborhood i ,y i The average value of the direction, x i ,y i Respectively represent the coordinate values of the i-th point cloud in the x direction and the y direction; Cov(x,y) represents the x, y covariance matrix; S22, with the vehicle-mounted laser scanning system trajectory point t i As the center, establish a local search area ROI and calculate the initial trajectory point t within the ROI area i and the end trajectory point t j The direction vector v i , calculate v in the ROI area i and road boundary point p 1 、p 2 、p 3 …p n main direction The angle θ 1 ,θ 2 ,θ 3 …θ n , this angle is called the angle characteristic value of the road boundary point, which satisfies the following formula: S23, using the vehicle-mounted laser scanning system trajectory point t i As the center, establish local ROI point by track point i , local ROI statistics i The eigenvalue of the angle at the inner boundary point θ i The cumulative sum Calculate point by point Traverse all trajectory points in turn and calculate the ROI corresponding to all trajectory points i The cumulative sum of the angle eigenvalues of the road boundary points within the region Intersection area ROI i Cumulative sum of angle eigenvalues Much larger than the sum of the eigenvalues of the angles in the straight line segment area Set the threshold θ threshold Separate straight line segments from intersection regions: Greater than θ threshold , then the current ROI i The area is a point cloud of intersections, otherwise it is a point cloud of straight line segments, which satisfies the following formula:
2. According to claim 1, a road scene type recognition method based on vehicle-mounted laser point cloud data, It is characterized in that The step S1 specifically includes the following steps: S11. Based on the vehicle-mounted laser point cloud data, the ground point cloud is obtained by using the cloth simulation filtering method; S12, based on the ground point cloud obtained in step S11, the ground point cloud is segmented using a supervoxel method, each point of the supervoxel has the same semantic label, and the road boundary point clouds with similar normals are segmented into the same semantic label, and all the road boundary point clouds in the label form a voxel block; S13, calculate the average angle Δθ between the normal vector of the point cloud in each voxel block of the ground point cloud supervoxel and the Z axis of the point cloud data coordinate system, calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block, and calculate the difference Δh between the maximum elevation and the minimum elevation of the point cloud in each voxel block according to the voxel block normal threshold t normal and the voxel block elevation difference threshold t height , calculate to satisfy t normal ,t height The extracted voxel block is the voxel block containing the road boundary point cloud; S14, based on the trajectory of the vehicle-mounted laser scanning system, retaining data points within a set range on both sides of the road, and removing data points outside the set range; S15, cluster the roughly extracted road boundary voxel blocks using Euclidean clustering, and set the cluster point number threshold t num , remove the clusters with points less than the threshold number and obtain the denoised voxel blocks, that is, the road boundary point cloud.
3. According to the method for identifying road scene types based on vehicle-mounted laser point cloud data in claim 1, It is characterized in that The step S3 specifically comprises the following steps: S31, based on the intersection point cloud extracted in S23, cluster the intersection point cloud into independent objects based on the connected branch clustering algorithm, where the cluster distance is C cluster , Based on the clustered independent objects at the intersection, the DGCNN network is used to classify the intersection type; S32. The intersection object is sampled into n points using the farthest point and input into the DGCNN network. The data format is n×3, where 3 represents the initial number of features in the feature dimension of the point cloud. The input data is calculated through 3 layers of Edge-Conv and 1 layer of multi-layer perceptron MLP to obtain n×1024 high-dimensional features of the sampling points. The semantically similar points are aggregated to extract local feature information at different levels of the point cloud. Finally, the maximum pooling is used to obtain n×1 global features, and the extracted global features are assigned to each point. The features of different levels extracted by each EdgeConv are spliced. All feature information is fused through MLP and the high-dimensional features are reduced in dimensionality. Finally, 2×1 is output to represent the probability value corresponding to the current object being classified into 2 categories by the network, so as to determine the category of the intersection object and realize the classification of crossroads and T-junctions.
4. A road scene type recognition system based on vehicle-mounted laser point cloud data, It is characterized in that The method comprises a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, the method steps as described in any one of claims 1 to 3 can be implemented.
