A method for automatic lane line extraction

Abstract

The present invention discloses a method for automatically extracting lane lines, which relates to the technical field of high-precision map production for autonomous driving. First, the pixel coordinates of the lane lines are extracted from a two-dimensional image. Second, the data is converted into lane point coordinates. Then, the map coordinate system of the lane points is calculated based on the posture data. Then, all lane points in an area are distance clustered using the DBSCAN algorithm, and noise points are filtered to obtain preliminary lane line data. Finally, the lane lines are merged, padded and interrupted, and complete lane line data is finally output. Lane line information is extracted based on the image data, lidar data and posture data collected by the unmanned vehicle, and is used for the production of high-precision map data. With the help of the lidar data, the two-dimensional lane pixel points extracted from the image can be solved into a three-dimensional map coordinate system with high conversion accuracy. In addition, the extracted multi-frame data are fused and post-processed, which can greatly increase the integrity of the lane lines and improve the automation rate of lane line extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision map production for autonomous driving, and in particular to a method for automatically extracting lane lines. Background Art

[0002] Lane markings are the left and right edges that define the range of a lane. Generally, lane ranges are determined by lane markings printed on the ground. These are roughly divided into five types: single dashed line, single solid line, double solid line, double dashed line, and dashed-solid line. They ensure that vehicles stay in the correct lane and provide safety for vehicle driving. They are the most important road element in the production of high-precision maps for autonomous driving. However, existing extraction methods suffer from low accuracy and low automation rates. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a method for automatically extracting lane lines, comprising the following steps:

[0004] S1. Obtain the image data of the current frame and extract the pixel coordinates of the lane lines from the image based on the deep learning lane detection model LaneATT.

[0005] S2. Obtain the lidar data for the current frame, select multiple rings that are close to the vehicle and can scan the ground, project the point cloud into the image's pixel coordinate system, calculate the intersection of the rings with the lane lines, and reverse-calculate the intersection points into the lidar's 3D coordinate system using a line segment intersection algorithm.

[0006] S3. Based on the current pose information, the 3D points of the extracted lanes are converted to the map coordinate system;

[0007] S4. Use the DBSCAN algorithm to perform distance clustering on all lane points extracted in an area and filter out noise points to obtain preliminary lane line data;

[0008] S5. Use the lateral extension algorithm and longitudinal distance judgment method to merge, fill in, and interrupt multiple discrete lane lines on the same road section, and finally output complete lane line data.

[0009] The technical solution further defined in the present invention is:

[0010] Furthermore, the line segment intersection algorithm in step S2 includes the following steps:

[0011] A1. Calculate the intersection point p0 between the lane line and the LiDAR scan line in the pixel coordinate system.

[0012] A2, record the point cloud coordinates p1 and p2 before and after the intersection;

[0013] A3, record the proportional coefficient I of the distance between the intersection point and the point cloud;

[0014] A4. Linearly interpolate the three-dimensional coordinate p of the intersection point based on the proportional coefficient I and the two lidar points p1 and p2 before and after the intersection point.

[0015] In the aforementioned automatic lane line extraction method, the conversion method in step S3 is set to calculate the current time posture information through linear interpolation, convert the point cloud coordinate system into the vehicle's baselink coordinate system through matrix transformation, and then convert the baselink coordinate system into the map coordinate system.

[0016] In the aforementioned automatic lane line extraction method, the noise filtering method in step S4 includes the following steps:

[0017] B1. Filter out shorter lane lines using a distance threshold.

[0018] B2. Calculate the angle between each point and its adjacent points, excluding the first and last points.

[0019] B3. Filter out salient points using the angle threshold to obtain preliminary lane line data.

[0020] In the aforementioned automatic lane line extraction method, the lane line merging method in step S5 is set to calculate the slope based on the first and last points of the lane line; then, based on the slope, the first and last points are extended outward by a certain lateral distance dx; then, the longitudinal distance dy from the extended first and last points to other lane lines is calculated. When the longitudinal distance is less than a threshold, the two lane lines are connected end to end (the two lane lines are considered to be very close).

[0021] In the aforementioned method for automatically extracting lane lines, the lane line interruption method in step S5 is set to read the extracted stop line data, traverse whether the lane line intersects with the stop line, and if so, interrupt the lane line to remove the part that exceeds the stop line.

[0022] In the aforementioned method for automatic lane line extraction, the lane line completion method in step S5 is set to read the extracted stop line data, traverse the lane line to see if it intersects with the stop line, and if not, extend the lane line horizontally by a certain distance, and then determine again whether it intersects with the stop line. If so, extend the lane line to the intersection point.

[0023] The beneficial effects of the present invention are:

[0024] This method can extract lane line information based on the image data, lidar data and posture data collected by the unmanned vehicle, and can be used to produce high-precision map data.

[0025] With the help of lidar data, the two-dimensional lane pixel points extracted from the image can be converted to a three-dimensional map coordinate system with high conversion accuracy. In addition, the fusion and post-processing of the extracted multi-frame data can greatly increase the integrity of the lane lines and improve the automation rate of lane line extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Schematic diagram of the lane point conversion from 2D to 3D according to the present invention;

[0027] Figure 2 Schematic diagram of lane line merging according to the present invention;

[0028] Figure 3 Schematic diagram of lane line filling and interruption according to the present invention. DETAILED DESCRIPTION

[0029] This embodiment provides a method for automatically extracting lane lines, the structure of which is as follows: Figures 1 to 3 As shown, the following steps are included

[0030] S1. Obtain the image data of the current frame and extract the pixel coordinates of the lane lines from the image based on the deep learning lane detection model LaneATT.

[0031] S2. Obtain the lidar data of the current frame, filter out multiple rings that are close to the vehicle and can scan the ground, and convert the point cloud coordinates to the pixel coordinate system of the image through the pre-calibrated exterior orientation elements, such as Figure 1 As shown in the figure, the intersection point p0 of the lane line and the lidar scan line in the pixel coordinate system is calculated through the line segment intersection algorithm; the point cloud coordinates p1 and p2 before and after the intersection are recorded; the proportional coefficient I of the distance between the intersection point and the point cloud before and after is recorded; the three-dimensional coordinate p of the intersection point is linearly interpolated based on the proportional coefficient I and the two lidar points p1 and p2 before and after the intersection;

[0032] S3. Obtain similar pose information based on the timestamp of the current lidar data, calculate the pose information at the current time through linear interpolation, convert the point cloud coordinate system to the vehicle's baselink coordinate system through matrix transformation, and then convert the baselink coordinate system to the map coordinate system;

[0033] S4. Use the DBSCAN algorithm to perform distance clustering on all lane points extracted within a region. Then, apply a distance threshold to the clustered results to filter out shorter lane lines. Then, calculate the angles between each point and its adjacent points (except the first and last points) and apply an angle threshold to filter out salient points, thereby obtaining preliminary lane line data.

[0034] S5. If the lane lines after clustering are discontinuous in the same road section or exceed the road section range, first calculate the slope based on the first and last points of the lane line, and then extend the first and last points outward by a certain distance dx (lateral distance) according to the slope; Figure 2 As shown, the distance dy (longitudinal distance) between the two points after extension and other lane lines is calculated. If the longitudinal distance is less than the threshold, the two lane lines are considered to be very close and the two lane lines are connected end to end. The extracted stop line data is read to check whether the lane line intersects with the stop line. If so, the lane line is interrupted and the part beyond the stop line is removed. Figure 3 Otherwise, the lane line is extended horizontally for a certain distance and it is judged again whether it intersects with the stop line. If it intersects, the lane line is extended to the intersection point. After merging, filling, interrupting and other operations, the complete lane line data is finally output.

[0035] This method can extract lane line information based on the image data, lidar data and posture data collected by the unmanned vehicle, and can be used to produce high-precision map data.

[0036] First, based on the deep learning model LaneATT, the pixel coordinates of the lane lines are extracted from the two-dimensional image. Secondly, the lidar data at the same time is converted to the image coordinate system and multiple rings are selected to intersect with the lane lines. The laser points before and after the intersection are obtained and the three-dimensional coordinates of the intersection are interpolated according to the ratio, which are the coordinates of the lane point. The map coordinate system of the lane point is calculated based on the posture data.

[0037] The DBSCAN algorithm is used to perform distance clustering on all lane points in an area, and noise points are filtered by distance and angle to obtain preliminary lane line data.

[0038] Finally, a lateral extension algorithm and longitudinal distance judgment are used to merge, fill in and interrupt multiple discrete lane lines on the same road section, and finally the complete lane line data is output.

[0039] With the help of lidar data, the two-dimensional lane pixel points extracted from the image can be converted to a three-dimensional map coordinate system with high conversion accuracy. In addition, the fusion and post-processing of the extracted multi-frame data can greatly increase the integrity of the lane lines and improve the automation rate of lane line extraction.

[0040] In addition to the above embodiments, the present invention may also have other implementations. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A method for automatically extracting lane lines, characterized by: The following steps are included S1. Obtain the image data of the current frame and extract the pixel coordinates of the lane lines from the image based on the deep learning lane detection model LaneATT. S2. Obtain the lidar data for the current frame, select multiple rings that are close to the vehicle and can scan the ground, project the point cloud into the image's pixel coordinate system, calculate the intersection of the rings with the lane lines, and reverse-calculate the intersection points into the lidar's 3D coordinate system using a line segment intersection algorithm. S3. Based on the current pose information, the 3D points of the extracted lanes are converted to the map coordinate system; S4. Use the DBSCAN algorithm to perform distance clustering on all lane points extracted in an area and filter out noise points to obtain preliminary lane line data; S5. Use the lateral extension algorithm and longitudinal distance judgment method to merge, fill in, and interrupt multiple discrete lane lines on the same road section, and finally output complete lane line data.

2. The method for automatically extracting lane lines according to claim 1, characterized in that: The line segment intersection algorithm in step S2 includes the following steps A1. Calculate the intersection point p0 between the lane line and the LiDAR scan line in the pixel coordinate system. A2, record the point cloud coordinates p1 and p2 before and after the intersection; A3, record the proportional coefficient I of the distance between the intersection point and the point cloud; A4. Linearly interpolate the three-dimensional coordinate p of the intersection point based on the proportional coefficient I and the two lidar points p1 and p2 before and after the intersection point.

3. The method for automatically extracting lane lines according to claim 1, wherein: The conversion method in step S3 is set to calculate the current time posture information through linear interpolation, convert the point cloud coordinate system into the vehicle's baselink coordinate system through matrix transformation, and then convert the baselink coordinate system into the map coordinate system.

4. The method for automatically extracting lane lines according to claim 1, wherein: The method for filtering noise in step S4 includes the following steps B1. Filter out shorter lane lines using a distance threshold. B2. Calculate the angle between each point and its adjacent points, excluding the first and last points. B3. Filter out salient points using the angle threshold to obtain preliminary lane line data.

5. The method for automatically extracting lane lines according to claim 1, wherein: The lane line merging method in step S5 is configured as follows: calculating a slope based on the first and last points of the lane line; then extending the first and last points outward by a certain lateral distance dx based on the slope; then calculating a longitudinal distance dy from the extended first and last points to other lane lines; and when the longitudinal distance is less than a threshold, connecting the two lane lines end to end.

6. The method for automatically extracting lane lines according to claim 1, characterized in that: The lane line interruption method in step S5 is set to read the extracted stop line data, traverse the lane line to see if it intersects with the stop line, and if so, interrupt the lane line to remove the part that exceeds the stop line.

7. The method for automatically extracting lane lines according to claim 1, characterized in that: The lane line filling method in step S5 is set to read the extracted stop line data, traverse to see whether the lane line intersects with the stop line, if not, extend the lane line horizontally by a certain distance, and again determine whether it intersects with the stop line, if so, extend the lane line to the intersection point.

Images