Unmanned vehicle high-precision map road network generation system and method

Abstract

The present invention provides a system and method for generating a high-precision road network map for an unmanned vehicle. The system includes a data receiving module, a semantic map module, a vectorization module, and a topological map module connected in sequence. The present invention makes full use of the precise trajectory information provided by the unmanned vehicle positioning module, as well as information such as laser point clouds and images, and develops a complete tool chain for semantic point cloud generation, vectorization extraction, and topological road network generation. It can quickly and automatically generate lane-level high-precision road network maps, meeting the core requirements of unmanned vehicle planning control and auxiliary positioning.

Description

Technical Field

[0001] The present invention relates to the field of high-precision maps for unmanned vehicles, and in particular to a system and method for generating a high-precision map road network for unmanned vehicles. Background Art

[0002] With the rapid rise of the autonomous driving industry, the exploration and application of unmanned vehicles in various industries have been greatly expanded. However, in actual implementation scenarios, the requirements for the intelligence and safety of individual vehicles are very high, and there are still great risks and challenges, which require continuous technical exploration and iteration. High-precision maps for autonomous driving can provide a wealth of prior information for individual vehicles, helping them to better perceive, locate, plan and control, and improve the intelligence and safety level of individual vehicles.

[0003] With the support of high-precision maps and intelligent networking technology, cars can also move from single-vehicle intelligence to vehicle network intelligence. High-precision maps can be accurate to the lane-level element granularity, and lanes can contain rich traffic rules and geometric features. A complete spatial topological relationship is built between lanes, and information on lane lines, ramps and other dimensions of the road is fed back to the car's electronic control system, so that the car always has "data to rely on" during autonomous driving, and performs beyond-visual-range perception, breaking through the limitations of the body sensor hardware, thereby achieving true high-level autonomous driving.

[0004] The high-precision road network map for unmanned vehicles uses the sensor data and trajectory positioning information of the unmanned vehicle itself to quickly build global lane-level topological map data, which is directly applied to the decision-making planning and auxiliary positioning of the unmanned vehicle. Therefore, the rapid construction of high-precision road network maps is currently a hot topic in the field of unmanned driving research.

[0005] At present, there are two main ways to collect raw data for high-precision maps. One is to use data collection vehicles with professional sensors, and the other is to use collection vehicles equipped with low-cost sensors to perform multi-batch crowdsourcing mapping. The equipment cost of professional collection vehicles is high, the data collection efficiency is low, and the map collection vehicles are not related to unmanned vehicles. Crowdsourcing collection uses low-cost sensors on the vehicle side to collect data. Due to the low accuracy of the data source, it is difficult to form a single batch of mapping needs. It is necessary to collect a large amount of data uploaded by the vehicle side and process it to generate maps with relatively reliable accuracy and integrity. The data collection vehicle is also a manned vehicle. This article uses the sensor equipment of commercial unmanned vehicles. The equipment cost is between the above two methods. In addition, the unmanned vehicle is equipped with complete high-precision sensor equipment and central data processors. It can be processed through the complete algorithm tool chain of this patent to quickly build lane-level high-precision road network maps in a single batch. The above-mentioned sensor equipment refers to laser radar, camera, IMU, GNSS and other equipment. Summary of the invention

[0006] In order to solve the problems of the prior art, the present invention provides a system and method for generating a high-precision road network map for an unmanned vehicle, which fully utilizes the precise trajectory information provided by the unmanned vehicle positioning module, as well as information such as laser point cloud and image, and formulates a complete tool chain for semantic point cloud generation, vectorization extraction, and topological road network generation, which can quickly and automatically generate lane-level high-precision road network maps, meeting the core requirements of unmanned vehicle planning, control, and auxiliary positioning.

[0007] The present invention provides a high-precision map road network generation system for unmanned vehicles, comprising a data receiving module, a semantic map module, a vectorization module, and a topological map module connected in sequence;

[0008] A data receiving module is used to receive the trajectory information and laser point cloud map sent by the unmanned vehicle positioning module, and receive the semantic segmentation image sent by the image processing module;

[0009] The semantic map module is used to perform road surface point cloud segmentation and line marking point cloud segmentation based on the input data and generate a point cloud map with semantic labels;

[0010] The vectorization module is used to extract vector skeleton lines from the point cloud data segmented from the semantic map and generate line strings containing semantic category attributes;

[0011] The topological map module is used to construct lanes for vector markings, complete virtual lanes, build the spatial inclusion relationship between markings and lanes, and the adjacency relationship between lanes, to form a global and complete lane network map.

[0012] The present invention also provides a method for generating a high-precision map road network for an unmanned vehicle, comprising the following steps:

[0013] 1) Using the data receiving module, receive the trajectory information and laser point cloud map sent by the unmanned vehicle positioning module, and receive the semantic segmentation image sent by the image processing module;

[0014] 2) Preprocessing the input data using a semantic map generation module to perform road surface segmentation and line marking segmentation of the laser point cloud;

[0015] 3) Using the marking vectorization generation module, the segmented marking point cloud is segmented longitudinally, the point cloud blocks in each segment are clustered, the key points and main directions of each cluster are extracted, the key points are used to connect the skeleton lines, and the skeleton lines are fitted and optimized to generate high-precision smooth vector marking lines;

[0016] 4) Using the topological map generation module, the mapping area is calculated and the global map grid is constructed. The grid is activated by the trajectory, and the situation of multiple trajectories in the same section is clustered. The undirected graph connected domain is constructed according to the information of the trajectory grid unit. The intersection area and non-intersection area are judged according to the single connectivity and multi-connectivity characteristics of the undirected graph nodes. The non-intersection single connectivity domain is longitudinally segmented and the topological relationship of the lateral lanes within the segment is constructed. The same-name lanes are matched and connected between the longitudinal segments, and the virtual lanes are constructed at the lane change points. The virtual lanes in the intersection area are constructed and connected according to the lane turning information. Finally, a global lane-level topological road network high-precision map is generated.

[0017] In step 1), the trajectory information includes time information, three-dimensional coordinate information, posture information, and position translation information of each frame of data at the time of image acquisition;

[0018] The laser point cloud map refers to the complete global point cloud data spliced ​​together using all frames of laser radar data;

[0019] The semantic segmentation image refers to the segmentation of image targets for the original color image data collected by the image acquisition device, and the output of a mask image with category labels.

[0020] In step 2), the preprocessing of the input data refers to segmenting and orderly organizing the laser point cloud map, semantic image and trajectory data;

[0021] The input data refers to trajectory data, laser point cloud data, and semantic image data; the ordered organization refers to segmenting and sorting according to the time series of the trajectory, and assigning a unique ID to each segment. Within each segment, each image sequence and corresponding trajectory are also sorted according to the time series and assigned an ID name; the segment length takes into account the actual road conditions.

[0022] The specific implementation method of the road surface segmentation of the segmented point cloud is to perform a denoising process of a bilateral filter on the segmented point cloud, perform uniform sampling and interpolation on the segmented point cloud, use the ground point cloud near the segmented track of the unmanned vehicle as the initial road surface seed patch, perform patch diffusion according to the normal vector consistency, intensity consistency, elevation consistency, and density consistency of the patch, and extract the road surface point cloud; process the point cloud of each segment in turn to complete the road surface segmentation of the global map;

[0023] The specific implementation method of the road marking segmentation is to project the segmented segmented road point cloud onto the semantic image plane, perform multi-frame voting using the relationship between the projected point cloud and the road marking semantic block, perform probability statistics according to the number of hits in the voting, mark the point cloud with a semantic label that meets the voting probability requirements, and then cluster the point cloud blocks in the segment using Euclidean distance, perform point cloud intensity mean statistics for each cluster, compare the reflection intensity prior value of the road marking point cloud with the statistical intensity mean, filter the clusters with relatively large intensity deviations, and thus screen out misclassified point cloud clusters; process the point cloud blocks of each segment in turn to complete the extraction of the road marking point cloud of the global map.

[0024] In step 3), the longitudinal differential segmentation of the marking point cloud refers to the segmentation according to the direction of the trajectory and at smaller intervals;

[0025] The extraction of key points refers to clustering the marking point cloud in the segment by Euclidean distance, and calculating the centroid of the point cloud blocks in the same cluster as the key point of the cluster;

[0026] The extraction of the main direction refers to calculating the direction vector with the largest variance for each cluster of point cloud blocks by using the PCA method as the main direction of the cluster;

[0027] The connection of the skeleton lines refers to the connection of key points according to the temporal sequence, spatial correlation and main direction consistency of the clusters;

[0028] The specific implementation method of the skeleton line connection is to create different vector line container lists (lane_list_1, lane_list_2, ..., lane_list_n) in the first segment according to the key points (v1, v2, ..., vn) extracted in the previous step, and push the corresponding key points; enter the next segment, check whether the cluster m of the current segment overlaps with the cluster n of the previous segment, and if there is an overlap, push the key point vm of the cluster into the corresponding vector line list lane_list_n, and complete The connection of the key point vm of cluster m this time; if there is no overlap, but the distance and main direction between cluster m and cluster n are less than the set threshold, the key point vm is pushed into the corresponding vector line list lane_list_n to complete the connection; if there is no overlap between cluster m in this segment and all clusters in the previous segment, and the distance threshold or main direction angle threshold exceeds the limit, the key point of the cluster is considered to be a new vector line vertex, a new vector line vertex list lane_list_n+1 is created, and the key point vm is pushed in; and so on, the connection of the key points of all segments is completed;

[0029] The fitting and optimization of the skeleton line refers to fitting the skeleton line according to the quadratic B-spline curve function, and simplifying the fitted curve using the Douglas Peuker method.

[0030] In step 4), the grid division of the mapping area refers to creating a regular grid using the bounding box of the mapping area as the grid range and the actual road width as the grid size;

[0031] The activation of the grid cells by using the driving trajectory refers to traversing the trajectory information to mark the cells where the trajectory points fall into the grid. If there are multiple trajectories in the same road section, the trajectories need to be clustered or de-redundanted first to ensure that a complete trajectory is retained in the same road section.

[0032] The construction of an undirected graph using a trajectory grid refers to defining an undirected graph G=(V, E), calculating the trajectory clustering vertices Vi in the grid for the activated grid cells, constructing edge connections Ei for the vertices of the adjacent domain grid cells, and generating an adjacency matrix of the undirected graph;

[0033] The undirected graph determines the single connected road section and the intersection road section, which means determining whether the same vertex Vi is connected by more than two edges according to the relationship between the vertices and edges of the constructed undirected graph. If there are more than two edge connections, the vertex is considered to be an intersection vertex, otherwise the vertex is considered to be a vertex in the single connected road section area;

[0034] The construction of the connection relationship of the intersection according to the undirected graph refers to determining the topological connection relationship between the edge connections Ei1 and Ei2, Ei1 and Ei3, Ei2 and Ei3, ... according to the topological relationship between the intersection vertex Vi and the multiple edges (Ei1, Ei2, Ei3, ...);

[0035] The connection relationship of the simply connected road segment constructed according to the undirected graph refers to determining the relationship between the edge connections of the simply connected domain road segment (Ej1, Ej2, Ej3, ..., Ejn) according to the undirected graph vertex Vj and the two adjacent edges Ej1 and Ej2 before and after, and the vertex Vj+1 and the two adjacent edges Ej2 and Ej3 before and after;

[0036] The specific implementation method of the longitudinal segmentation of the simply connected section Si is as follows: first, a maximum longitudinal segmentation threshold DL is set; from the starting position of the simply connected section, the change point of the vector marking line is detected along the direction of travel of the road; if a change point is detected, longitudinal cutting is immediately started for segmentation to generate longitudinal segments; if no change point is detected, cutting is started at the maximum longitudinal segmentation threshold DL to generate segments; wherein i represents the number of the simply connected domain, and m represents the segment number within the simply connected domain.

[0037] The change point of the vector marking line refers to the beginning and end points of the vector marking line, the intersection point of multiple vector marking lines, and the change point between the dotted line and the solid line of the same marking line;

[0038] The constructing of the transverse topological relationship of the vector markings in the segment refers to horizontally sorting the multiple vector markings in the segment, determining the baseline reference line of the markings by using the prior knowledge of the semantic labels of the markings, using the reference markings as the dividing line of the opposite lanes, constructing the lane surface for the two adjacent markings on the left and right sides of the reference line according to the sorted markings, and sorting the constructed lane surface lane on the left and right sides with the baseline reference line as the interval, starting from 1, 2, ... on the left side, and starting from -1, -2, ... on the right side;

[0039] The matching and connection of lane planes between segments refers to aligning the joints between segments according to the relationship between the adjacent segment reference lines and the nearest neighbor; in the case where multiple lanes correspond to one lane or one lane corresponds to multiple lanes, a virtual line segment is constructed according to the principle that the steering angle is not less than the angle threshold to complete the longitudinal connection of the lane surface;

[0040] The construction and connection of virtual markings in the intersection area refers to constructing the connection relationship of the intersection according to an undirected graph, mapping the simply connected segment and the connection relationship through the positional relationship between the grid unit of the undirected graph edge and the longitudinal segment section, matching the lanes in the two segments according to the driving rules, and connecting the left and right side markings of the two matched segmented lanes according to the Bezier curve to generate a virtual lane surface of the intersection.

[0041] The present invention also provides a system terminal for generating a high-precision map road network for an unmanned vehicle, comprising a processor, a memory, a communication interface, and a computer program; the computer program is stored in the memory and is configured to be executed by the processor, and the computer program includes instructions for executing a method for generating a high-precision map road network for an unmanned vehicle.

[0042] The present invention also provides a computer program for a method for generating a high-precision map road network for an unmanned vehicle, and the computer program includes instructions for executing the method for generating a high-precision map road network for an unmanned vehicle.

[0043] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores the above-mentioned computer program.

[0044] The beneficial effects of the present invention are:

[0045] 1. Receive the trajectory information and laser point cloud map output by the unmanned vehicle positioning module, and receive the semantic segmentation image output by the image processing module;

[0046] 2. Use the idea of ​​regional segmentation, division and merging to fuse laser point cloud and semantic image to quickly extract semantic point cloud of markings and road surface;

[0047] 3. Use longitudinal differential segmentation, key point extraction, skeleton line tracking and skeleton line optimization to quickly extract vector markings;

[0048] 4. Use trajectory gridding, grid-based undirected graph generation and connected domain judgment to construct road-level topological relationships, identify intersection areas and single-connected areas based on road-level topological relationships; segment the markings longitudinally in the single-connected sections, construct transverse topological relationships and lanes for the markings in the segments, match and connect the segments according to the connectivity of the markings; construct virtual lanes in the intersection area according to lane turning traffic rules, and connect them to the single-connected sections.

[0049] 5. The method in this paper proposes three complete tool chains: semantic map generation, feature vectorization and topological map construction. The proposed lane construction and connection technology for intersection sections and single-connected sections under the guidance of an undirected graph connectivity domain can quickly generate high-precision vector topological maps, significantly improving the accuracy and efficiency of high-precision road network map generation, making the entire high-precision road network map generation system have good accuracy, timeliness and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. It includes a data receiving module, a semantic map module, a vectorization module, and a topology map module connected in sequence;

[0051] Figure 1 Schematic diagram of the operation process of the semantic map module;

[0052] Figure 2 Schematic diagram of the vectorization module operation process;

[0053] Figure 3 This is a schematic diagram of the operation process of the topology map module;

[0054] Figure 4 This is a schematic diagram of the module of the high-precision map road network generation system for unmanned vehicles;

[0055] Figure 5 A schematic diagram of the terminal of the high-precision map network generation system for unmanned vehicles;

[0056] Figure 6 A schematic diagram for matching and connecting lane planes between segments;

[0057] Figure 7 To connect according to Bezier curves, a schematic diagram of the virtual lane surface of the intersection is generated;

[0058] Figure 8 It is a semantic point cloud map;

[0059] Fig. 9 For vectorized line maps;

[0060] Fig.10 A topological map. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] In a first aspect, the embodiments of the present application provide a system and device for generating a high-precision map road network for an unmanned vehicle, the method comprising the following steps:

[0063] 1. The data receiving module receives the trajectory information and laser point cloud map sent by the unmanned vehicle positioning module, and receives the semantic segmentation image sent by the image processing module.

[0064] The trajectory information includes time information, three-dimensional coordinate information, posture information, and position translation information of each frame of data at the time of image acquisition.

[0065] The laser point cloud map refers to the complete global point cloud data spliced ​​together using all frames of laser radar data.

[0066] The semantic segmentation image refers to the segmentation of image targets for the original color image data collected by the image acquisition device, and the output of a mask image with category labels.

[0067] 2. A semantic map generation module pre-processes the input data and performs road segmentation and line segmentation of the laser point cloud, such as Figure 1 shown.

[0068] 2-1 The preprocessing of the input data refers to segmenting and orderly organizing the laser point cloud map, semantic image and trajectory data.

[0069] The input data refers to trajectory data, laser point cloud data, and semantic image data; ordered organization refers to segmenting and sorting according to the time series of the trajectory, and assigning a unique ID to each segment. Within each segment, each image sequence and corresponding trajectory are also sorted according to the time series and assigned an ID name; the segment length takes into account the actual road conditions, for example, segmentation is performed according to a threshold of 20 meters.

[0070] 2-2 The specific implementation method of the road segmentation of the segmented point cloud is to perform bilateral filtering noise reduction on the segmented point cloud, perform uniform sampling and interpolation on the segmented point cloud, use the ground point cloud near the segmented track of the unmanned vehicle as the initial road seed patch, perform patch diffusion according to the normal vector consistency, intensity consistency, elevation consistency, and density consistency of the patch, and extract the road point cloud. The point cloud of each segment is processed in turn to complete the road segmentation of the global map.

[0071] The specific implementation method of the line segmentation described in 2-3 is to project the segmented road surface point cloud onto the semantic image plane, perform multi-frame voting based on the relationship between the projected point cloud and the semantic block of the line, perform probability statistics based on the number of hits in the voting, mark the point cloud with semantic labels that meet the voting probability requirements, and then cluster the point cloud blocks within the segment based on the Euclidean distance, perform statistics on the mean value of the point cloud intensity for each cluster, compare the reflection intensity prior value of the line point cloud with the statistical mean value of the intensity, filter the clusters with relatively large intensity deviations, and thus screen out the misclassified point cloud clusters. Process the point cloud blocks of each segment in turn to complete the extraction of the line point cloud of the global map.

[0072] 3. Marking vector generation module, such as Figure 2 As shown, the segmented marking point cloud is segmented longitudinally, the point cloud blocks in each segment are clustered, the key points and main directions of each cluster are extracted, the key points are used to connect the skeleton lines, and the skeleton lines are fitted and optimized to generate high-precision smooth vector marking lines.

[0073] 3-1 The longitudinal differential segmentation of the marking point cloud refers to longitudinal cutting according to the direction of trajectory travel and at smaller intervals, such as a distance of 1 meter, with a certain degree of overlap between adjacent segmentation units, for example, setting an overlap of 30% of the segment length.

[0074] The extraction of key points described in 3-2 refers to clustering the marking point cloud within the segment based on the Euclidean distance, and calculating the centroid points of the point cloud blocks in the same cluster as the key points of the cluster.

[0075] The extraction of the main direction refers to calculating the direction vector with the largest variance for each cluster of point cloud blocks by using the PCA method as the main direction of the cluster.

[0076] The connection of the skeleton lines described in 3-3 refers to the connection of key points according to the temporal sequence, spatial correlation and main direction consistency of the clusters. The specific implementation method is to create different vector line container lists (lane_list_1, lane_list_2, ..., lane_list_n) within the first segment according to the key points (v1, v2, ..., vn) extracted in the previous step, and push the corresponding key points; enter the next segment, check whether the cluster m of the current segment overlaps with the cluster n of the previous segment, if there is an overlap, push the key point vm of the cluster into the corresponding vector line list lane_list_n, and complete the connection of the key point vm of this cluster m; if there is no overlap, but the distance and main direction of the cluster m and the cluster n are less than the set threshold, then push the key point vm into the corresponding vector line list lane_list_n to complete the connection; if there is no overlap between the cluster m in the current segment and all the clusters of the previous segment, and the distance threshold or the main direction angle threshold exceeds the limit, then the key point of the cluster is considered to be a new vector line vertex, a new vector line vertex list lane_list_n+1 is created, and the key point vm is pushed in. And so on, complete the connection of the key points of all the cut segments.

[0077] The fitting and optimization of the skeleton line described in 3-4 refers to fitting the skeleton line according to the quadratic B-spline curve function and simplifying the fitted curve using the Douglas Peuker method.

[0078] 4. Use the topological map generation module, such as Figure 3 As shown in the figure, the mapping area is calculated and the global map grid is constructed, and the grid is activated by the trajectory. The situation of multiple trajectories in the same section is clustered, and the undirected graph connected domain is constructed according to the information of the trajectory grid unit. The intersection area and non-intersection area are judged according to the single connectivity and multi-connectivity characteristics of the undirected graph nodes. The non-intersection single connectivity domain is longitudinally segmented and the topological relationship of the lateral lanes in the segment is constructed. The same-name lanes are matched and connected between the longitudinal segments, and the virtual lanes of the lane change points are constructed. The virtual lanes in the intersection area are constructed and connected according to the lane turning information, and finally a global lane-level topological road network high-precision map is generated.

[0079] 4-1 The grid division of the mapping area mentioned above refers to creating a regular grid using the bounding box of the mapping area as the grid range and the actual road width as the grid size.

[0080] 4-2 Using driving trajectories to activate grid cells means traversing the trajectory information to mark the cells where the trajectory points fall into the grid. If there are multiple trajectories on the same road section, the trajectories need to be clustered or de-redundanted first to ensure that a complete trajectory is retained on the same road section.

[0081] 4-3 The use of trajectory grids to construct an undirected graph refers to defining an undirected graph G = (V, E), calculating the trajectory clustering vertices Vi within the grid for activated grid cells, building edge connections Ei for the vertices of adjacent domain grid cells, and generating an adjacency matrix of the undirected graph.

[0082] 4-4 The undirected graph described in claim 4-4 determines whether a single connected section and an intersection section refers to determining whether there are more than two edges connecting the same vertex Vi based on the relationship between the vertices and edges of the constructed undirected graph. If there are more than two edges connecting the same vertex Vi, the vertex is considered to be an intersection vertex, otherwise the vertex is considered to be a vertex in the single connected section area.

[0083] 4-5 The construction of the connection relationship of the intersection according to the undirected graph refers to determining the topological connection relationship between the edge connections Ei1 and Ei2, Ei1 and Ei3, Ei2 and Ei3, ... according to the topological relationship between the intersection vertex Vi and multiple edges (Ei1, Ei2, Ei3, ...).

[0084] The connection relationship of constructing a simply connected segment based on an undirected graph as described in 4-6 refers to determining the relationship between the edge connections of the simply connected domain segment (Ej1, Ej2, Ej3, ..., Ejn) based on the undirected graph vertex Vj and the two adjacent edges Ej1 and Ej2 before and after, and the vertex Vj+1 and the two adjacent edges Ej2 and Ej3 before and after.

[0085] 4-7 The segmentation of the longitudinal segments of the simply connected road section Si described above is specifically implemented as follows: first, a maximum longitudinal segmentation threshold DL is set, and from the starting position of the simply connected road section, the change point of the vector marking line is detected along the direction of travel of the road. If a change point is detected, the longitudinal cutting is immediately started to segment and generate longitudinal segments. If no change point is detected, the cutting is started at the maximum longitudinal segmentation threshold DL to generate segments. Where i represents the number of the simply connected domain, and m represents the segment number in the simply connected domain.

[0086] 4-8 The changing points of vector markings refer to the starting and ending points of vector markings, the intersection points of multiple vector markings, the changing points between the dotted line and the solid line of the same marking, etc.

[0087] 4-9 The construction of a lateral topological relationship for the vector markings within the segment refers to horizontally sorting multiple vector markings within the segment, using the prior knowledge of the semantic labels of the markings to determine the baseline reference line, using the reference markings as the dividing line of the opposite lanes, and constructing a lane surface for two adjacent markings on the left and right sides of the reference line according to the sorted markings, and sorting the constructed lane surface lane on the left and right sides with the baseline reference line as the interval, starting from 1, 2, ... on the left side, and starting from -1, -2, ... on the right side.

[0088] The lane plane matching and connection between segments as described in 4-10 refers to aligning the joints between segments according to the reference lines of adjacent segments and the nearest neighbor relationship. In the case where there are multiple lanes corresponding to one lane or one lane corresponding to multiple lanes, a virtual line segment is constructed according to the principle that the steering angle is not less than the angle threshold (for example, 120 degrees) to complete the longitudinal connection of the lane surface. Figure 6 As shown, the lane surface with ID -2 in segment section1 corresponds to the lane surfaces with IDs -2 and -3 in segment section2. Section 1 (-2) is directly connected to section 2 (-2). Section 1 (-2) and section 2 (-3) need to be constructed as follows Figure 6 The virtual marking line shown completes the connection between section 1 (-2) and section 2 (-3).

[0089] The construction and connection of virtual markings in the intersection area described in 4-11 refers to mapping the simply connected segments and connectivity relationships through the positional relationship between the grid units of the undirected graph edges and the longitudinal sections according to the edge connection relationship described in 4-5, matching the lanes in the two segments according to the driving rules (for example, the left turn lane on the right side of the current segment reference line is connected to the right lane of the left turn segment, and the right turn lane is connected to the right lane of the right turn segment according to the right lane of the current segment reference line). For the two matched lane segments, the left and right lane markings are connected according to the Bezier curve to generate a virtual lane surface for the intersection. The Bezier curve is not limited to a specific order in the present invention and can be flexibly selected according to actual conditions. Figure 7 As shown, the construction principle of the second-order Bezier curve is used for public display and schematic diagram explanation.

[0090] P i =(1-t) 2 P a +2t(1-t)P e +t 2 P c i=1,2,...,m

[0091] P j =(1-t) 2 P b +2t(1-t)P f +t 2 P d j=1,2,...,n

[0092] Among them, t∈[0,1], e and f are the intersection points of the extended lines of the left marking (ac) and the right marking (bd) matched by the segmented lane, P i and Pj are the corresponding points i and j on the left and right markings of the virtual lane. The number of vertices of the virtual markings can be calculated by dividing the value range of t into m equal parts and n equal parts respectively.

[0093] In a second aspect, the present application embodiment provides a system for generating a high-precision map road network for an unmanned vehicle, such as Figure 4 As shown, including:

[0094] A data receiving module is used to receive the trajectory information and laser point cloud map sent by the unmanned vehicle positioning module, and receive the semantic segmentation image sent by the image processing module;

[0095] The semantic map module is used to perform road surface point cloud segmentation and line marking point cloud segmentation based on the input data and generate a point cloud map with semantic labels, such as Figure 8 As shown;

[0096] The vectorization module is used to extract the vector skeleton line from the point cloud data segmented from the semantic map and generate a line string containing semantic category attributes. The resulting vectorized line map is as follows: Fig. 9 As shown;

[0097] The topological map module is used to construct lanes for vector markings, complete virtual lanes, build the spatial inclusion relationship between markings and lanes, and the adjacency relationship between lanes, to form a global and complete lane network map, such as Fig.10 shown.

[0098] In a third aspect, an embodiment of the present application provides a terminal, such as Figure 5 As shown, it includes: a processor, a memory, a communication interface, and a computer program;

[0099] The computer program is stored in the memory and is configured to be executed by the processor, wherein the computer program includes instructions for executing the method as described in the first aspect.

[0100] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program enables a server to execute the method described in the first aspect.

[0101] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising computer instructions, wherein the computer instructions are executed by a processor according to the method described in the first aspect.

[0102] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, the above is only a preferred embodiment of the present invention. Since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with the technical field is within the technical scope disclosed by the present invention. For ordinary technicians in the technical field, changes or replacements that can be easily thought of should be covered within the protection scope of the present invention without departing from the principle of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for generating a high-precision map road network for an unmanned vehicle, characterized in that: The following steps are involved: 1) Use the data receiving module to receive the trajectory information and laser point cloud map sent by the unmanned vehicle positioning module, and receive the semantic segmentation image sent by the image processing module; 2) Using the semantic map generation module, the input data is preprocessed to perform road surface segmentation and road marking segmentation of the laser point cloud; 3) Using the marking vectorization generation module, the segmented marking point cloud is segmented longitudinally, the point cloud blocks in each segment are clustered, the key points and main directions of each cluster are extracted, the key points are used to connect the skeleton lines, and the skeleton lines are fitted and optimized to generate high-precision smooth vector marking lines; The longitudinal differential segmentation of the marking point cloud refers to the segmentation in the direction of the trajectory at smaller intervals; The extraction of key points refers to clustering the marking point cloud in the segment by Euclidean distance, and calculating the centroid of the point cloud blocks in the same cluster as the key point of the cluster; The extraction of the main direction refers to calculating the direction vector with the largest variance as the main direction of each cluster of point cloud blocks by using the PCA method; The connection of the skeleton line refers to the connection of key points according to the temporal sequence, spatial correlation and main direction consistency of the cluster; the specific implementation method of the connection of the skeleton line is to create different vector line container lists (lane_list_1, lane_list_2, ..., lane_list_n) in the first segment according to the key points (v1, v2, ..., vn) extracted in the previous step, and press the corresponding key points; enter the next segment, check whether the cluster m of the current segment overlaps with the cluster n of the previous segment, and if there is an overlap, press the key point vm of the cluster into the corresponding vector line container list. The list of vector marking lines lane_list_n is used to complete the connection of the key point vm of this cluster m; if there is no overlap, but the distance and main direction of the cluster m and cluster n are less than the set threshold, the key point vm is pushed into the corresponding vector marking line list lane_list_n to complete the connection; if the cluster m in this segment has no overlap with all the clusters in the previous segment, and the distance threshold or the main direction angle threshold exceeds the limit, the key point of the cluster is considered to be a new vector marking line vertex, and a new vector marking line vertex list lane_list_n+1 is created, and the key point vm is pushed in; and so on, the connection of the key points of all segments is completed; The fitting and optimization of the skeleton line refers to fitting the skeleton line according to the quadratic B-spline curve function and simplifying the fitted curve using Douglas Peuker's method; 4) Using the topological map generation module, the mapping area is calculated and the global map grid is constructed. The grid is activated by the trajectory, and multiple trajectories in the same section are clustered. The undirected graph connected domain is constructed according to the information of the trajectory grid unit. The intersection area and non-intersection area are judged according to the single connectivity and multi-connectivity characteristics of the undirected graph nodes. The non-intersection single connectivity domain is longitudinally segmented and the topological relationship of the lateral lanes within the segment is constructed. The same-name lanes are matched and connected between the longitudinal segments, and the virtual lanes are constructed at the lane change points. The virtual lanes in the intersection area are constructed and connected according to the lane turning information. Finally, a global lane-level topological road network high-precision map is generated.

2. The method for generating a high-precision map road network for an unmanned vehicle according to claim 1, characterized in that: In step 1), the trajectory information includes time information, three-dimensional coordinate information, posture information, and position translation information of each frame of data at the time of image acquisition; The laser point cloud map refers to the complete global point cloud data spliced ​​together using all frames of laser radar data; The semantic segmentation image refers to the segmentation of image targets for the original color image data collected by the image acquisition device, and the output of a mask image with category labels.

3. The method for generating a high-precision map road network for an unmanned vehicle according to claim 1, characterized in that: In step 2), the preprocessing of the input data refers to segmenting and orderly organizing the laser point cloud map, semantic image and trajectory data; The input data refers to trajectory data, laser point cloud data, and semantic image data; the ordered organization refers to segmenting and sorting according to the time series of the trajectory, and assigning a unique ID to each segment. Within each segment, each image sequence and corresponding trajectory are also sorted according to the time series and assigned an ID name; the segment length takes into account the actual road conditions.

4. The method for generating a high-precision map road network for an unmanned vehicle according to claim 1 or 3, characterized in that: In step 2), the specific implementation method of the road segmentation of the segmented point cloud is to perform bilateral filtering noise reduction processing on the segmented point cloud, perform uniform sampling and interpolation on the segmented point cloud, use the ground point cloud near the segmented trajectory of the unmanned vehicle as the initial road seed patch, perform patch diffusion according to the normal vector consistency, intensity consistency, elevation consistency, and density consistency of the patch, and extract the road point cloud; process the point cloud of each segment in turn to complete the road segmentation of the global map; The specific implementation method of the road marking segmentation is to project the segmented segmented road point cloud onto the semantic image plane, perform multi-frame voting using the relationship between the projected point cloud and the road marking semantic block, perform probability statistics according to the number of hits in the voting, mark the point cloud with a semantic label that meets the voting probability requirements, and then cluster the point cloud blocks in the segment using Euclidean distance, perform point cloud intensity mean statistics for each cluster, compare the reflection intensity prior value of the road marking point cloud with the statistical intensity mean, filter the clusters with relatively large intensity deviations, and thus screen out misclassified point cloud clusters; process the point cloud blocks of each segment in turn to complete the extraction of the road marking point cloud of the global map.

5. The method for generating a high-precision map road network for an unmanned vehicle according to claim 1, characterized in that: In step 4), the mapping area is divided into grids, which means using the bounding box of the mapping area as the grid range and the actual road width as the grid size to create a regular grid; Activating grid cells using driving trajectories means traversing the trajectory information to mark the cells where the trajectory points fall into the grid. If there are multiple trajectories on the same road section, the trajectories need to be clustered or de-redundanted first to ensure that a complete trajectory is retained on the same road section. Constructing an undirected graph using trajectory grids means defining an undirected graph G=(V,E), calculating the trajectory clustering vertices Vi within the grid for the activated grid cells, constructing edge connections Ei for the vertices of the adjacent domain grid cells, and generating an adjacency matrix of the undirected graph; The undirected graph determines the single connected road section and the intersection road section, which means determining whether the same vertex Vi is connected by more than two edges according to the relationship between the vertices and edges of the constructed undirected graph. If there are more than two edge connections, the vertex is considered to be an intersection vertex, otherwise the vertex is considered to be a vertex in the single connected road section area; Constructing the connection relationship of the intersection according to the undirected graph means determining the topological connection relationship between the edge connections Ei1 and Ei2, Ei1 and Ei3, Ei2 and Ei3, ... according to the topological relationship between the intersection vertex Vi and multiple edges (Ei1, Ei2, Ei3, ...); The connection relationship of the simply connected road segment constructed according to the undirected graph refers to determining the relationship between the edge connections of the simply connected domain road segment (Ej1, Ej2, Ej3, ..., Ejn) according to the undirected graph vertex Vj and the two adjacent edges Ej1 and Ej2 before and after, and the vertex Vj+1 and the two adjacent edges Ej2 and Ej3 before and after; The specific implementation method of the longitudinal segmentation of the simply connected road section Si is as follows: first, a maximum longitudinal segmentation threshold DL is set; from the starting position of the simply connected road section, a change point of the vector marking line is detected along the direction of travel of the road; if a change point is detected, longitudinal cutting is immediately started to segment to generate longitudinal segments; if no change point is detected, cutting is started at the maximum longitudinal segmentation threshold DL to generate segments; wherein i represents the number of the simply connected domain, and m represents the segment number in the simply connected domain; The change point of the vector marking line refers to the beginning and end points of the vector marking line, the intersection point of multiple vector marking lines, and the change point between the dotted line and the solid line of the same marking line; The constructing of the transverse topological relationship of the vector markings in the segment refers to horizontally sorting the multiple vector markings in the segment, determining the baseline reference line of the markings by using the prior knowledge of the semantic labels of the markings, using the reference markings as the dividing line of the opposite lanes, constructing the lane surface for the two adjacent markings on the left and right sides of the reference line according to the sorted markings, and sorting the constructed lane surface lane on the left and right sides with the baseline reference line as the interval, starting from 1, 2, ... on the left side, and starting from -1, -2, ... on the right side; The matching and connection of lane planes between segments refers to aligning the joints between segments according to the relationship between the adjacent segment reference lines and the nearest neighbor; in the case where multiple lanes correspond to one lane or one lane corresponds to multiple lanes, a virtual line segment is constructed according to the principle that the steering angle is not less than the angle threshold to complete the longitudinal connection of the lane surface; The construction and connection of virtual markings in the intersection area refers to the construction of the connection relationship of the intersection according to an undirected graph, mapping the simply connected segment and the connection relationship through the positional relationship between the grid unit of the undirected graph edge and the longitudinal segment section, matching the lanes in the two segments according to the driving rules, and connecting the left and right side markings of the two matched lanes according to the Bezier curve to generate a virtual lane surface of the intersection.

6. A high-precision map road network generation system terminal for unmanned vehicles, characterized by: The invention comprises a processor, a memory, a communication interface, and a computer program; the computer program is stored in the memory and is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to claim 1.

Images