A method for constructing high-precision maps in unstructured road scenes

Abstract

The present invention provides a method for constructing a high-precision map suitable for unstructured road scenes, and relates to the field of autonomous driving technology. The method uses a data collection vehicle that has been jointly calibrated with a camera and a laser radar to collect unstructured road data and generate a depth image, and produces a point cloud map based on the collected road data, selects a corresponding algorithm to extract the road area according to the road environment, annotates the generated point cloud map with road boundaries, determines the road width, and generates virtual lanes with speed and direction according to traffic rules, manually corrects the lane lines marked in step 4, manually annotates the positions of traffic lights and road signs, and saves the map in the required format. Based on road boundaries, the present invention proposes an effective method for constructing high-precision maps of unstructured roads in environments such as closed parks and towns and villages, ensuring that it can correctly assist unmanned vehicle perception, positioning, and planning in the current autonomous driving software framework.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method for constructing a high-precision map suitable for unstructured road scenarios. Background Art

[0002] As one of the core roles in the field of autonomous driving, high-precision maps can help cars perceive complex road information in advance, such as slope, curvature, and heading, and combine intelligent path planning to allow cars to make correct decisions. High-precision maps are digital maps with higher accuracy and more data dimensions than traditional electronic maps. The higher accuracy is reflected in the map's accuracy to the centimeter level, and the data dimension is reflected in the fact that in addition to road information, it also includes other information around the road related to traffic. It is generally believed that high-precision maps must be introduced to achieve the commercialization of autonomous driving above level L3. Its main service object is the autonomous driving system, that is, the on-board host. Compared with human drivers, the on-board host does not have the logical analysis ability of humans. With the help of high-precision maps, unmanned vehicles can be provided with a global view that sensors do not have, including road and traffic facility information outside the sensor monitoring range, helping unmanned vehicles to predict and avoid known obstacles in advance and improve the efficiency of overall control. However, due to the characteristics of high-precision maps requiring high accuracy and containing a lot of information, the production of high-precision maps is more difficult than ordinary electronic maps.

[0003] Since the development of autonomous driving, commercialization has been achieved for specific scenarios, but there is still a big gap from the highest level of autonomous driving under any drivable conditions. The current development direction of autonomous driving is mainly focused on urban roads and highways, which have clear lane lines and traffic signs. There is no good solution for unstructured roads without lane lines or obvious road boundaries. On structured roads, map manufacturers, car companies and Internet giants use static high-precision map production solutions that integrate multiple sensors such as lidar, cameras, GPS and IMU to ensure centimeter-level accuracy by identifying the locations of lane lines, road signs, and static traffic facilities such as road signs and traffic lights. There is a concept of lanes on structured roads, and the lanes are one-way, so the lane driving direction and speed information can be directly marked on the high-precision map. However, on unstructured roads, the concept of lanes is unclear or there are no lanes, so the solution for structured roads cannot be directly applied to unstructured road scenarios. In summary, building high-precision maps suitable for autonomous driving in unstructured road scenarios is a direction for the future development of autonomous driving to a higher level.

[0004] There are currently the following problems to be solved in the production of high-precision maps in unstructured road scenarios: (1) The concept of lanes is unclear. A road can be driven in both directions, and it is impossible to mark the lanes of the driving direction like structured roads; (2) Compared with structured roads, the perception of unstructured road scenarios is simpler overall, and there are fewer things to perceive. Therefore, the multi-sensor fusion solution of the map collection vehicle on structured roads has too much data redundancy in unstructured road scenarios. In addition, the sensors on structured roads are expensive, and data processing is difficult and costly; (3) The existing deep learning high-precision map production method has a high degree of automation, but relies on the previous road feature images for model training. In unstructured road scenarios, the irregular road boundaries and the lack of road signs make it difficult to produce labeled images, and it is difficult to train an unstructured road recognition model with a high recognition rate. In addition, deep learning requires a large amount of data and a large amount of data calculation, and the cost of map production is high. Summary of the invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a method for constructing high-precision maps in unstructured road scenarios, based on road boundaries, to ensure that it can correctly assist unmanned vehicle perception, positioning and planning in the current autonomous driving software framework.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A method for constructing a high-precision map in an unstructured road scene includes the following steps:

[0008] Step 1: Use a data collection vehicle with jointly calibrated cameras and lidars to collect unstructured road data, and use the lidar data image data to generate a depth image;

[0009] Step 2: Create a point cloud map based on the collected road data, and select the corresponding algorithm to extract the road area according to the road environment;

[0010] Step 3: Mark the road boundaries on the point cloud map generated in step 2;

[0011] Step 4: Determine the road width and generate a virtual lane with speed and direction according to traffic rules;

[0012] Step 5: Manually correct the lane lines marked in step 4, manually mark the locations of traffic lights and road signs, and save the map in the required format.

[0013] The step 1 comprises:

[0014] Step 1.1: Calibrate the camera to obtain the camera intrinsic parameter matrix K; after determining the installation positions of the camera and the lidar on the vehicle body, jointly calibrate the camera and the lidar to obtain the extrinsic parameter matrix, including the rotation matrix R and the translation vector t from the camera to the lidar;

[0015] Step 1.2: Drive the data collection vehicle to the right on the target road at a constant speed. The data collection vehicle is equipped with a computer device to save the collected data as a data package of a .bag file;

[0016] Step 1.3: Read the image data and lidar data in the data packet, project the lidar data onto the image corresponding to the timestamp, and generate a depth image.

[0017] The step 2 comprises:

[0018] Step 2.1: Play back the data collected by the collection vehicle, select the corresponding algorithm to generate the point cloud map, and determine the algorithm for generating the point cloud map based on the road environment;

[0019] Step 2.2: Sample the collected image data, find the vanishing point of the image through texture-based road vanishing point detection, and crop the image according to the vanishing point position to extract the area of ​​interest;

[0020] Step 2.3: Segment the road region of interest extracted in the previous step, cut out the road area and the non-road area, and extract the boundary points of the road area;

[0021] Step 2.4: Use formula (5) to transform the detected road boundary into the LiDAR 3D coordinate system, where z c is the z-axis coordinate value of the road boundary line in the camera coordinate system, (u,v) is the coordinate of the road boundary point in the two-dimensional image, (x L ,y L ,z L ) are the coordinates of the road boundary points in the laser radar coordinate system;

[0022]

[0023] Step 2.5: According to the installation positions of the laser radar and the vehicle body, the road boundary points in the laser radar coordinate system are converted to the vehicle body coordinate system, and then the road boundary points in the vehicle body coordinate system are converted to the world coordinate system. Here, it corresponds to the point cloud map and uses the starting point of the data collection vehicle as the origin;

[0024] Step 2.6: Save and output the coordinates of the road boundary points based on the world coordinate system;

[0025] Step 2.7: Repeat steps 2.2 to 2.6 above until the data processing is completed.

[0026] The step 3 comprises:

[0027] Step 3.1: traverse the road boundary points of the world coordinate system obtained in step 2, set the z-axis coordinates of the road boundary points to zero, and project each road boundary point to the xoy plane;

[0028] Step 3.2: Traverse the coordinate point set of the road boundary points and optimize the coordinate point set of the road boundary points. If the distance between the subsequently updated coordinate point and the existing coordinate point is less than 0.05m, delete the point with the earlier timestamp and retain the later updated road boundary point. Then fit the obtained road boundary points to the left and right road boundary curves.

[0029] Step 3.3: Sample the fitted left and right road boundary curves respectively, and save the coordinates of the points in csv format;

[0030] Step 3.4: Save the road boundary as a vector map format under the Autoware autonomous driving platform.

[0031] The step 4 comprises:

[0032] Step 4.1: Subtract the coordinates corresponding to the two road boundary curves obtained in step 3 to obtain the road width;

[0033] Step 4.2: If the road width is greater than twice the urban lane width, according to the traffic rule of driving on the right, generate two lane lines at 1 / 4 of the distance from the left and right road boundaries, save them in vector map format, and jump to step 4.4 to continue; otherwise, proceed to step 4.3;

[0034] Step 4.3: When the road width is less than twice the urban lane width, the road boundary is used as the reference to generate the left and right lane lines 1.35 meters away from the left and right road boundary lines respectively, and save them in vector map format.

[0035] Step 4.4: Read the generated lane line lane, judge the road bend and perform special processing; if the road bend is an arc, skip the special processing step and execute step 5; if the road corner is a special angle of non-arc boundary, the intersection of lane line lane1 before the bend and lane line lane2 after the bend is point Q, calculate the distance l from point Q to the intersection of the road boundary line, take point P with a distance l to point Q on the reverse extension line from point Q to the intersection of the road boundary line, and the arc with point P as the center is tangent to lane1 and lane2, remove the points in lane1 and lane2 outside the two tangent points, and retain the continuous arc lane line.

[0036] The step 5 comprises:

[0037] Step 5.1: Read the point cloud map generated in step 2, the road boundary generated in step 3, and the lane line generated in step 4 into the Unity software, and manually adjust the fit between the lane line and the road boundary line and the point cloud map according to the road inclination;

[0038] Step 5.2: Mark the specific locations of traffic lights and road signs in the point cloud map, and connect the traffic lights to their corresponding lanes;

[0039] Step 5.3: Export the manually adjusted map and save it in the desired format.

[0040] The beneficial effect of adopting the above technical solution is that: the method for constructing high-precision maps in unstructured road scenes provided by the present invention is based on road boundaries, and proposes an effective method for constructing high-precision maps of unstructured roads in environments such as closed parks and towns and villages, which can overcome the defect of lack of road signs on unstructured roads, and the generated high-precision maps can be compatible with existing autonomous driving software platforms. The present invention converts some manual annotations into automatic generation, reduces the proportion of manual links on the basis of ensuring the centimeter-level accuracy of high-precision maps, and improves the efficiency of constructing high-precision maps. After comparing the high-precision map production methods of structured roads, the characteristics and similarities and differences between unstructured roads and structured roads are analyzed, and the high-precision maps under unstructured roads are generated using the same format as under structured roads, ensuring that they can correctly assist unmanned vehicles in perception, positioning and planning in the current autonomous driving software framework. Compared with the high-precision map construction solution for roads in closed park environments in the prior art, the scope of application of the present invention is wider, it can adapt to a variety of road models, and the high-precision maps constructed are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flow chart of a method for constructing a high-precision map in an unstructured road scenario provided by an embodiment of the present invention;

[0042] Figure 2 An abstract schematic diagram for explaining image segmentation below a vanishing point provided by an embodiment of the present invention;

[0043] Figure 3 A virtual lane generation diagram when the road width is greater than 7 meters provided in an embodiment of the present invention;

[0044] Figure 4 A virtual lane generation diagram when the road width is insufficient provided by an embodiment of the present invention;

[0045] Figure 5 A schematic diagram of generating lane lines at a normal turning point provided by an embodiment of the present invention;

[0046] Figure 6 A schematic diagram of lane adjustment at a curve at a special angle provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] In the description of this embodiment, words indicating directions or sensor installation positions such as "front", "rear", "left" and "right" are merely examples of the present invention. They do not require that the device or sensor must be installed in a specific direction, and therefore should not be construed as limiting the scope of protection of the present invention.

[0049] In this embodiment, at least one computer device is set up for unstructured road data collection and calculation when generating point cloud maps. At the same time, the computer device needs to store perception information, positioning information and map information. The computer can be an industrial computer or a laptop computer. The vehicle body is equipped with a 16-line or more laser radar for collecting 3D information around the road, and a monocular camera for collecting road image information.

[0050] like Figure 1 As shown, the method for constructing a high-precision map in an unstructured road scenario applicable to this embodiment is described as follows.

[0051] Step 1: Collect unstructured road scene information; including:

[0052] Step 1.1: This embodiment uses an unstructured road data collection vehicle equipped with a 16-line laser radar and a monocular camera. Before use, the camera is calibrated to obtain the camera intrinsic parameter matrix K. After determining the installation position of the camera and the laser radar on the vehicle body, the camera and the laser radar are jointly calibrated to obtain the rotation matrix R and translation matrix t from the camera to the laser radar. In this embodiment, the sensor models used include: Velodyne VLP16 laser radar, FILR BFS-GPE-16S2C-CS monocular camera.

[0053] Step 1.2: The data collection vehicle drives at a constant speed on the right side of the target road. The vehicle is equipped with a computer device. The computer is equipped with Ubuntu 18.04 system and ROS Melodic. The collected data is stored in a .bag format data package. Considering the accuracy of the later point cloud map construction, the starting point and end point of each collection route are overlapped as much as possible.

[0054] Step 1.3: Read the image data and lidar data in the data packet, map the lidar data to the image corresponding to the timestamp, and generate a depth image.

[0055] Step 2: Create a point cloud map based on the collected road data, and extract the road area based on the road environment selection algorithm; including:

[0056] Step 2.1: Replay the data collected by the collection vehicle, subscribe to the lidar data to generate a point cloud map. In this embodiment, NDT is used to construct the point cloud map. SLAM algorithm can also be selected in actual operation. However, due to the large number of roads around the experiment in this example, the mapping effect of some SLAM algorithms based on feature extraction is not as good as NDT.

[0057] Step 2.2: Sample the collected image data. The image collected by the collection vehicle is 24fps, and the image data of 8fps is obtained by interval sampling. The sampling interval is determined by the driving speed of the collection vehicle. If the collection vehicle is driving slowly, multiple samplings can be performed to ensure the detection accuracy while minimizing unnecessary repeated calculations.

[0058] This embodiment uses texture-based road vanishing point detection. After the road vanishing point is determined by the local soft voting method, the parallel line where the vanishing point is located is used as the segmentation line, and the image below the segmentation line is retained for processing. Considering that the elements above the vanishing point are mostly irrelevant to the road, such as the sky and trees, Figure 2 As shown, unnecessary calculations can be reduced after clipping, and according to the sensor installation method of this embodiment, there are only a small number of laser radar scanning lines above the straight line where the vanishing point is located, and extracting the area of ​​interest can speed up the detection speed.

[0059] Step 2.3: Use the two-dimensional OTSU threshold segmentation algorithm to segment the road area of ​​interest extracted in the previous step to obtain the road area and other environmental areas, extract the largest connected area as the road area, delete the smaller connected areas, and extract the boundary points of the road area.

[0060] Step 2.4: Project the road boundary points extracted in the previous step to the LiDAR 3D coordinate system to obtain the road boundary points in the LiDAR 3D coordinate system; obtain the transformation formula from the parameters of the jointly calibrated camera and LiDAR:

[0061]

[0062] The coordinates of the road boundary point in the 3D camera coordinate system are P c =[x c ,y c ,z c ] -T When , the coordinate correspondence of the pixel coordinate point in the laser radar coordinate system can be transformed as shown in formula (2), where E is the external parameter matrix obtained by the joint calibration of the laser radar and camera.

[0063]

[0064] The coordinates of the road boundary point P in the camera coordinate system c =[x c ,y c ,z c ] -T , connect the optical center to the projection point P in the physical coordinate system I =(x I ,y I ), then the coordinate transformation relationship between the two points is:

[0065]

[0066] Among them, f x and f y is the camera focal length.

[0067] The coordinates of the road boundary point P in the camera coordinate system c =[x c ,y c ,z c ] -T And the camera intrinsic parameter matrix K, and the coordinate point P in the image coordinate system img =[u,v], the transformation relationship is shown in formula (4):

[0068]

[0069] Among them, (u0, v0) is the optical axis projection coordinate in the image pixel coordinate system, that is, the principal point coordinate.

[0070] From the above formula, we can see that only z c The point is unknown. Combining the above formula and the coordinates (x,y) corresponding to the road boundary coordinates (u,v) in the 2D image, we can calculate the depth image coordinates (x c ,y c ,z c ) to obtain z c The value of the point is obtained by using formula (5) to obtain the coordinates of the road boundary point relative to the laser radar coordinate system and save them.

[0071]

[0072] Step 2.5: Read the positioning data of the car at this time to obtain the displacement of the car relative to the starting position. Convert the road boundary points based on the laser radar coordinate system obtained in step 2.4 to the vehicle coordinate system according to the installation position of the laser radar relative to the vehicle body. Then, convert it to the world coordinate system based on the initial position of the data collection vehicle as the origin to obtain the boundary coordinate points of the unstructured road in the real environment.

[0073] This embodiment selects GPS+IMU to perform real-time positioning of the vehicle. The characteristics of GPS data are real-time update and no cumulative error, but it will jump when disturbed or blocked; while the data of IMU relies on its own measurement and is not interfered by the outside world, that is, the data will not jump due to external information. The disadvantage is that the data is obtained by integration, so there will be cumulative errors. After the data of the two are fused, both the cumulative error can be eliminated and the jump can be avoided.

[0074] However, if you are generating a high-precision map of a closed park, or if the GPS signal in the target unstructured road area is affected by the surrounding environment and is not suitable for use, you can use IMU alone for positioning. At this time, the coordinates of the road boundary points in the previous step based on the laser radar coordinate system are converted to coordinates relative to the origin (i.e. the starting point of the data collection vehicle).

[0075] Step 2.6: Save and output the road boundary points based on the world coordinate system.

[0076] Step 2.7: Repeat the above steps 2.2 to 2.6 until the data processing is completed;

[0077] After the above steps, the output is a set of unstructured road boundary points based on the world coordinate system. At this time, there will be many repeated road boundary points in the set, and the coordinate point set needs to be optimized, that is, the coordinate point set is traversed. If the distance between two coordinate points is less than 0.05m, only one of them will be retained to screen out the boundary points that are closer to the real world road.

[0078] Step 3: Annotate the road boundaries on the unstructured road point cloud map, including:

[0079] Step 3.1: Traverse the road boundary points of the world coordinate system obtained in step 2, and project the points to the xoy plane, that is, set the z-axis coordinate value of the road boundary point to zero, that is, regard the road surface as an ideal plane.

[0080] Step 3.2: Since there will be redundancy of duplicate points in adjacent sampling images in the previous step, if the coordinate point updated later is less than 0.05m away from the existing coordinate point, the point with the earlier timestamp will be deleted, and the road boundary point updated later will be retained to fit the left and right boundary curves of the road.

[0081] Step 3.3: Sample the two fitted road boundary curves and save the coordinates of the points in csv format.

[0082] Step 3.4: This embodiment uses the vector map format to output and test the high-precision map, and later uses the Autoware autonomous driving platform to test the feasibility of the generated high-precision map.

[0083] Step 4: Determine the road width and generate a virtual lane with speed and direction according to traffic rules, including:

[0084] Step 4.1: Calculate the road width by subtracting the coordinates of the left and right road boundary curves obtained in step 3, and select a method for generating virtual lanes in the subsequent sequence according to the road width.

[0085] Step 4.2: If the road width is greater than 7 meters (equivalent to twice the width of a city lane), this type of unstructured road can be regarded as an ordinary two-lane city road. At this time, according to the traffic rules of driving on the right and the road speed limit required by local regulations, two lanes (lane format in the vector map format) are generated at 1 / 4 of the distance from the left and right road boundaries, such as Figure 3 As shown, then jump to step 4.4 to continue; otherwise, go to step 4.3. Lane is a set of points where the vehicle can drive, including the driving direction, the maximum speed during driving, and the recommended driving speed. The speed unit is km / h.

[0086] Step 4.3: When the road width is less than 7 meters, the unstructured road cannot be simply segmented. Considering that most vehicles on such roads are small and medium-sized, and the vehicle width is generally 1.6m to 2.0m, the road boundary is used as the reference to generate two lanes on the left and right, 1.35 meters away from the road boundary, such as Figure 4 As shown. Lane is a collection of vector points, including the driving direction, maximum speed and recommended speed. The speed unit is km / h.

[0087] Step 4.4: Read the generated lanes, judge the road bends and perform special processing. There are two specific cases: The first is the normal road bends. The road itself will take the bends into consideration when designing, and the road boundary is an arc, such as Figure 5 As shown, the special processing steps are skipped at this time, and step 5 is executed directly; the second case is that the road corner is a special angle with a non-arc boundary, for example, the road corner is a right angle. At this time, the above steps will be discontinuous in lane 1 of the road before the bend and lane 2 after the bend, and there is an intersection Q. The distance l from point Q to the intersection of the road boundary is calculated, and point P with a distance l to point Q is taken on the reverse extension line from point Q to the intersection of the road boundary. The arc with point P as the center is tangent to lane 1 and lane 2, and the points in lane 1 and lane 2 outside the two tangent points are removed to retain the continuous lanes, as shown in FIG. Figure 6 shown.

[0088] Step 5: Manually correct the lane lines marked in step 4, manually mark the locations of traffic lights and road signs, and save the map in the required format, including:

[0089] Step 5.1: Read the point cloud map generated in step 2, the road boundary generated in step 3, and the lane generated in step 4 into the unity software. This embodiment uses the MapToolbox tool under unity. Unity is a software platform that supports 3D content editing and production. MapToolbox is an open source plug-in that can edit and adjust maps in VMF format. Since the lanes and road boundaries in the above steps are generated under ideal conditions (removing z-axis interference and simplifying calculations), the roads in the real environment are not strictly horizontal, so it is necessary to manually adjust the fit between the lanes and road boundaries and the point cloud map in this step.

[0090] Step 5.2: Mark the specific locations of traffic lights and road signs in the software, and connect the traffic lights to their corresponding lanes.

[0091] Step 5.3: Export the manually adjusted map and save it in the required format. In this embodiment, it is saved in the vector map format that can be used by the Autoware autonomous driving platform for subsequent verification.

[0092] In order to test the effectiveness of the present invention, the Autoware platform was used in a simulation environment and an actual environment to test whether the high-precision map constructed by the method of the present invention can be correctly read and used for positioning and planning.

[0093] In the test, the high-precision maps in the point cloud map and vector map formats obtained by the present invention can be correctly loaded by Autoware1.14 version, and the maps can be used to realize NDT-matching positioning, and the open-planner planning algorithm can be used for path planning on the generated lanes; during automatic driving, the direction and speed of the lane can be correctly read and sent to the wire-controlled chassis, and the vehicle can travel quickly at the recommended speed.

[0094] The present invention proposes a method for constructing a high-precision map of an unstructured road based on the road boundary, which can overcome the defect of the lack of road signs on unstructured roads, and proposes an effective method for constructing a high-precision map of unstructured roads in environments such as closed parks and towns and villages, and the generated high-precision map can be compatible with the existing autonomous driving software platform. Since unstructured roads have fewer elements to perceive during the construction process compared to structured roads, the present invention converts some manual annotations into automatic generation, which reduces the proportion of manual links on the basis of ensuring the centimeter-level accuracy of the high-precision map and improves the efficiency of high-precision map construction. Compared with the high-precision map construction solution for roads in closed park environments in the prior art, the present invention has a wider scope of application, can adapt to a variety of road models, and constructs a high-precision map with higher accuracy. The high-precision map constructed by the present invention has high accuracy and rich data, and can provide auxiliary positioning, auxiliary perception and auxiliary planning functions for autonomous driving vehicles.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for constructing a high-precision map in an unstructured road scene, characterized by: The method comprises the following steps: Step 1: Use a data collection vehicle with jointly calibrated cameras and lidars to collect unstructured road data, and use the lidar data image data to generate a depth image; Step 2: Create a point cloud map based on the collected road data, extract the road area based on the road environment selection algorithm, and extract the boundary points of the road area; Step 3: Annotate the road boundaries on the point cloud map generated in step 2, and fit the obtained road boundary points into two left and right road boundary curves; Step 4: Determine the road width and generate a virtual lane with speed and direction according to traffic rules; specifically: Step 4.1: Subtract the coordinates corresponding to the two road boundary curves obtained in step 3 to obtain the road width; Step 4.2: If the road width is greater than twice the urban lane width, according to the traffic rule of driving on the right, generate two lane lines at 1 / 4 of the distance from the left and right road boundaries, save them in vector map format, and jump to step 4.4 to continue; otherwise, proceed to step 4.3; Step 4.3: When the road width is less than twice the urban lane width, the road boundary is used as the reference to generate the left and right lane lines 1.35 meters away from the left and right road boundary lines respectively, and save them in vector map format; Step 4.4: Read the generated lane line lane, judge the road bend and perform special processing; if the road bend is an arc, skip the special processing step and execute step 5; if the road corner is a special angle of non-arc boundary, the intersection of lane line lane1 before the bend and lane line lane2 after the bend is point Q, and calculate the distance from point Q to the intersection of the road boundary line l , the distance from point Q to the point Q on the reverse extension line of the intersection of the road boundary line is l Point P is the point where the arc with point P as the center is tangent to lane1 and lane2. The points in lane1 and lane2 outside the two tangent points are removed to retain the continuous arc lane line. Step 5: Manually correct the lane lines marked in step 4, manually mark the locations of traffic lights and road signs, and save the map in the required format.

2. The method for constructing a high-precision map in an unstructured road scene according to claim 1, characterized in that: The step 1 comprises: Step 1.1: Calibrate the camera to obtain the camera intrinsic parameter matrix K; after determining the installation positions of the camera and the laser radar on the vehicle body, jointly calibrate the camera and the laser radar to obtain the extrinsic parameter matrix, including the rotation matrix R and the translation matrix t from the camera to the laser radar; Step 1.2: Drive the data collection vehicle to the right on the target road at a constant speed. The data collection vehicle is equipped with a computer device to save the collected data as a data package of a .bag file; Step 1.3: Read the image data and lidar data in the data packet, project the lidar data onto the image corresponding to the timestamp, and generate a depth image.

3. The method for constructing a high-precision map in an unstructured road scene according to claim 2, characterized in that: The step 2 comprises: Step 2.1: Play back the data collected by the collection vehicle, select the corresponding algorithm to generate the point cloud map, and determine the algorithm for generating the point cloud map based on the road environment; Step 2.2: Sample the collected image data, find the vanishing point of the image through texture-based road vanishing point detection, and crop the image according to the vanishing point position to extract the area of ​​interest; Step 2.3: Segment the road region of interest extracted in the previous step, cut out the road area and the non-road area, and extract the boundary points of the road area; Step 2.4: The detected road boundary is transformed into the LiDAR 3D coordinate system by using formula (5), where z c is the z-axis coordinate value of the road boundary line in the camera coordinate system, ( u , v ) is the coordinate of the road boundary point in the two-dimensional image, ( x L , y L , z L ) are the coordinates of the road boundary points in the laser radar coordinate system; (5) Step 2.5: According to the installation positions of the laser radar and the vehicle body, the road boundary points in the laser radar coordinate system are converted to the vehicle body coordinate system, and then the road boundary points in the vehicle body coordinate system are converted to the world coordinate system. Here, it corresponds to the point cloud map and uses the starting point of the data collection vehicle as the origin; Step 2.6: Save and output the coordinates of the road boundary points based on the world coordinate system; Step 2.7: Repeat steps 2.2 to 2.6 above until the data processing is completed.

4. The method for constructing a high-precision map in an unstructured road scene according to claim 3, characterized in that: In the step 2.3, a two-dimensional OTSU threshold segmentation algorithm is used to segment the road area and the non-road area, and the largest connected area is extracted as the road area. After deleting the smaller connected areas, the boundary points of the road area are extracted.

5. The method for constructing a high-precision map in an unstructured road scene according to claim 3, characterized in that: The step 3 comprises: Step 3.1: traverse the road boundary points of the world coordinate system obtained in step 2, set the z-axis coordinates of the road boundary points to zero, and project each road boundary point to the xoy plane; Step 3.2: Traverse the coordinate point set of the road boundary points and optimize the coordinate point set of the road boundary points. If the distance between the subsequently updated coordinate point and the existing coordinate point is less than 0.05m, delete the point with the earlier timestamp and retain the later updated road boundary point. Then fit the obtained road boundary points to the left and right road boundary curves. Step 3.3: Sample the fitted left and right road boundary curves respectively, and save the coordinates of the points in csv format; Step 3.4: Save the road boundary as a vector map format under the Autoware autonomous driving platform.

6. The method for constructing a high-precision map in an unstructured road scene according to claim 1, characterized in that: In step 4.2 and step 4.3, lane is a set of points where the vehicle can travel, including the driving direction, the maximum speed during driving, and the recommended driving speed. The speed unit is km / h .

7. The method for constructing a high-precision map in an unstructured road scene according to claim 1, characterized in that: The step 5 comprises: Step 5.1: Read the point cloud map generated in step 2, the road boundary generated in step 3, and the lane line generated in step 4 into the Unity software, and manually adjust the fit between the lane line and the road boundary line and the point cloud map according to the road inclination; Step 5.2: Mark the specific locations of traffic lights and road signs in the point cloud map, and connect the traffic lights to their corresponding lanes; Step 5.3: Export the manually adjusted map and save it in the desired format.

Images