Method for identifying large-curvature lane line

Abstract

The invention discloses a method for identifying a large-curvature lane line. The method comprises the following steps of: constructing a pavement 3D point cloud containing RGB (Red, Green, Blue) information; performing visual transformation on the pavement 3D point cloud with the RGB color to generate ground 2D images under different visual angles; lane line extraction is carried out on the ground 2D images under different visual angles; and in the same world coordinate system, generating a final lane line by using a point cloud clustering algorithm. According to the method, the accuracy of large-curvature lane line identification can be improved, and the dependence on a high-precision map is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent driving, and more specifically, to a method for recognizing lane lines with large curvature. Background Art

[0002] Lane recognition is crucial in autonomous driving, helping the system monitor and track lane markings in real time, ensuring the vehicle stays centered in the lane, thereby improving driving safety. Accurate lane recognition helps the system prevent lane departures, assists with vehicle decisions such as lane changes and overtaking, and collaborates with other autonomous driving features to optimize the driving experience and overall safety. However, in reality, lanes with large curvature (i.e., steeply curved lanes) present a significant challenge for intelligent driving systems. This is primarily because sensors may have difficulty capturing the changing curvature of the lanes, and data processing algorithms may also experience errors due to the complex shape of the lanes. Therefore, accurately and quickly recognizing lanes with large curvature is crucial to improving recognition capabilities and system stability. Summary of the Invention

[0003] The present invention provides a method for recognizing lane lines with large curvature, which solves the problem of inaccurate recognition of lane lines with large curvature by existing intelligent driving vehicles. It can improve the vehicle's recognition ability, increase the accuracy of lane line recognition with large curvature, and reduce dependence on high-precision maps.

[0004] To achieve the above objectives, the present invention provides the following technical solutions:

[0005] A method for recognizing lane lines with large curvature, comprising:

[0006] Construct a 3D point cloud of the road surface containing RGB information;

[0007] Perform visual transformation on the 3D point cloud of the road surface with RGB colors to generate 2D images of the ground under different perspectives;

[0008] Extract lane lines from ground 2D images at different perspectives;

[0009] In the same world coordinate system, the point cloud clustering algorithm is used to generate the final lane lines.

[0010] Preferably, constructing a 3D point cloud of a road surface containing RGB information includes:

[0011] Acquire a 2D image of the road surface in front of the vehicle;

[0012] Obtain road surface 3D point cloud data through LiDAR, and extract the road surface 3D point cloud from the 3D point cloud data to filter out non-road parts;

[0013] The road surface 2D image and the road surface 3D point cloud are registered to map the colors of the road surface 2D image to the points of the road surface 3D point cloud, and the corresponding mapping relationship is stored.

[0014] Preferably, the lane line extraction from the ground 2D image at different viewing angles includes:

[0015] A deep learning algorithm is used to extract lane lines from multi-view ground 2D images to obtain an array of coordinate values ​​for each 2D lane line.

[0016] Preferably, generating the final lane line using a point cloud clustering algorithm in the same world coordinate system includes:

[0017] Converting the 2D lane line coordinate value array into a 3D lane line coordinate value array according to the mapping relationship;

[0018] The point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by the mapping to obtain the final synthesized 3D lane line.

[0019] Preferably, registering the road surface 2D image and the road surface 3D point cloud to map the colors of the road surface 2D image to points of the road surface 3D point cloud includes:

[0020] Calculate the 2D image coordinates corresponding to each 3D point;

[0021] Obtaining color information corresponding to 2D image coordinates from the ground 2D image using a linear interpolation method of an RGB image;

[0022] The obtained color information is combined with the coordinates of the road surface 3D point cloud to obtain a point cloud containing color information.

[0023] Preferably, calculating the 2D image coordinates corresponding to each 3D point includes:

[0024] The internal and external parameters of the camera and lidar are used to perform coordinate conversion to obtain the corresponding 2D image coordinates.

[0025] Preferably, the RANSAC algorithm is used to extract the road surface 3D point cloud from the 3D point cloud data.

[0026] Preferably, a deep learning algorithm such as LaneNet, CLRNet or LaneATT is used to extract lane lines from multi-view ground 2D images.

[0027] Preferably, the DBSCAN point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by mapping.

[0028] This invention provides a method for identifying lane lines with large curvature. By combining laser point cloud information with a 3D road surface point cloud containing RGB information, the method then uses perspective transformation to generate ground images from different perspectives. Lane lines are extracted from these images, and finally, clustering is used to generate the final lane lines within a unified world coordinate system. This method addresses the inaccurate recognition of lane lines with large curvature by existing intelligent driving vehicles, improving vehicle recognition capabilities and accuracy while reducing reliance on high-precision maps. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments.

[0030] Figure 1 It is a schematic diagram of a method for identifying lane lines with large curvature provided by the present invention.

[0031] Figure 2 The figure is a flow chart of a method for identifying lane lines with large curvature provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to enable those skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and implementation methods.

[0033] In response to the problem that current intelligent driving vehicles have inaccurate recognition of lane lines with large curvature, the present invention provides a method for recognizing lane lines with large curvature, which solves the problem that existing intelligent driving vehicles have inaccurate recognition of lane lines with large curvature. It can improve the vehicle's recognition ability, increase the accuracy of recognizing lane lines with large curvature, and reduce dependence on high-precision maps.

[0034] like Figure 1 and Figure 2 As shown, a method for recognizing lane lines with large curvature includes:

[0035] S1: Construct a 3D point cloud of the road surface containing RGB information;

[0036] S2: Perform visual transformation on the 3D point cloud of the road surface with RGB colors to generate 2D images of the ground under different perspectives;

[0037] S3: Extract lane lines from ground 2D images at different viewing angles;

[0038] S4: In the same world coordinate system, the point cloud clustering algorithm is used to generate the final lane lines.

[0039] Furthermore, constructing a 3D point cloud of the road surface containing RGB information includes:

[0040] Acquire a 2D image of the road surface in front of the vehicle;

[0041] Obtain road surface 3D point cloud data through LiDAR, and extract the road surface 3D point cloud from the 3D point cloud data to filter out non-road parts;

[0042] The road surface 2D image and the road surface 3D point cloud are registered to map the colors of the road surface 2D image to the points of the road surface 3D point cloud, and the corresponding mapping relationship is stored.

[0043] Specifically, a camera captures a 2D image of the road ahead of the vehicle, and a lidar radar acquires 3D point cloud data. Non-road sections are removed to create a 3D point cloud. A 3D point cloud is constructed based on the 2D image and 3D point cloud. To map image colors to points in the point cloud, registration of the 2D image and 3D point cloud is required. After acquiring the 3D point cloud with RGB color information, 2D ground images are generated from different positions and perspectives to create 2D ground images from different perspectives. After generating the 2D ground images from different perspectives, a 2D lane line deep learning algorithm is used to extract lane lines. The extracted 2D lane line coordinates are mapped back to a 3D world coordinate system using the image-point cloud mapping relationship, generating a multi-perspective lane line point cloud. Finally, a point cloud clustering algorithm is used to generate the final detection results. This method effectively addresses the problem of inaccurate or overfitting model recognition of lane lines with large curvature. By fusing lane line detection results from different perspectives using the world coordinate system, the detection rate and recognition accuracy of lane lines with large curvature are improved.

[0044] Furthermore, the lane line extraction from the ground 2D image at different viewing angles includes:

[0045] A deep learning algorithm is used to extract lane lines from multi-view ground 2D images to obtain an array of coordinate values ​​for each 2D lane line.

[0046] Furthermore, in the same world coordinate system, generating the final lane line using the point cloud clustering algorithm includes:

[0047] Converting the 2D lane line coordinate value array into a 3D lane line coordinate value array according to the mapping relationship;

[0048] The point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by the mapping to obtain the final synthesized 3D lane line.

[0049] Furthermore, registering the road surface 2D image and the road surface 3D point cloud to map the colors of the road surface 2D image to points of the road surface 3D point cloud includes:

[0050] Calculate the 2D image coordinates corresponding to each 3D point;

[0051] Obtaining color information corresponding to 2D image coordinates from the ground 2D image using a linear interpolation method of an RGB image;

[0052] The obtained color information is combined with the coordinates of the road surface 3D point cloud to obtain a point cloud containing color information.

[0053] In practical applications, linear interpolation of images is a commonly used image processing technique used to estimate new pixel values ​​between known pixel values. The purpose is to estimate the target pixel value based on the surrounding known pixel values, thereby generating an image with a smooth transition.

[0054] Furthermore, the calculation of the 2D image coordinates corresponding to each 3D point includes:

[0055] The internal and external parameters of the camera and lidar are used to perform coordinate conversion to obtain the corresponding 2D image coordinates.

[0056] In practical applications, after acquiring a 3D point cloud of the road surface with RGB color information, the extrinsic parameter matrices R and t in the formula for converting point cloud coordinates to image coordinates are adjusted. Simulating camera rotation or translation generates 2D images of the ground surface from different perspectives. This process requires preserving the mapping relationship [x, y, z] <-> [u, v] between the original point cloud and image coordinates. After generating 2D images of the ground surface from different perspectives, a 2D lane line deep learning algorithm is used to extract the lane lines, obtaining an array of 2D coordinate values ​​[u, v] for each lane line.

[0057] Furthermore, the RANSAC algorithm is used to extract the road surface 3D point cloud from the 3D point cloud data.

[0058] Furthermore, the deep learning algorithm of LaneNet, CLRNet or LaneATT is used to extract lane lines from multi-view ground 2D images.

[0059] Furthermore, the DBSCAN point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by mapping.

[0060] The present invention provides a method for identifying lane lines with large curvature. This method combines laser point cloud information to construct a three-dimensional point cloud of the road surface containing RGB information. This method then uses perspective transformation to generate ground images from different perspectives. Lane lines are extracted from these images, and finally, clustering is used within a unified world coordinate system to generate the final lane lines. This method addresses the inaccurate recognition of lane lines with large curvature by existing intelligent driving vehicles, improving vehicle recognition capabilities and accuracy, while reducing reliance on high-precision maps.

[0061] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for identifying lane lines with large curvature, characterized in that: include: Construct a 3D point cloud of the road surface containing RGB information; Perform visual transformation on the 3D point cloud of the road surface with RGB colors to generate 2D images of the ground under different perspectives; Extract lane lines from ground 2D images at different perspectives; In the same world coordinate system, the point cloud clustering algorithm is used to generate the final lane lines.

2. The method for recognizing lane lines with large curvature according to claim 1, characterized in that: The construction of a road surface 3D point cloud containing RGB information includes: Acquire a 2D image of the road surface in front of the vehicle; Obtain road surface 3D point cloud data through LiDAR, and extract the road surface 3D point cloud from the 3D point cloud data to filter out non-road parts; The road surface 2D image and the road surface 3D point cloud are registered to map the colors of the road surface 2D image to the points of the road surface 3D point cloud, and the corresponding mapping relationship is stored.

3. The method for recognizing lane lines with large curvature according to claim 2, characterized in that: The lane line extraction from the ground 2D image at different viewing angles includes: A deep learning algorithm is used to extract lane lines from multi-view ground 2D images to obtain an array of coordinate values ​​for each 2D lane line.

4. The method for recognizing lane lines with large curvature according to claim 3, characterized in that: In the same world coordinate system, the point cloud clustering algorithm is used to generate the final lane line, including: Converting the 2D lane line coordinate value array into a 3D lane line coordinate value array according to the mapping relationship; The point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by the mapping to obtain the final synthesized 3D lane line.

5. The method for recognizing lane lines with large curvature according to claim 4, characterized in that: The registering the road surface 2D image and the road surface 3D point cloud to map the color of the road surface 2D image to the points of the road surface 3D point cloud includes: Calculate the 2D image coordinates corresponding to each 3D point; Obtaining color information corresponding to 2D image coordinates from the ground 2D image using a linear interpolation method of an RGB image; The obtained color information is combined with the coordinates of the road surface 3D point cloud to obtain a point cloud containing color information.

6. The method for recognizing lane lines with large curvature according to claim 5, characterized in that: Calculating the 2D image coordinates corresponding to each 3D point includes: The internal and external parameters of the camera and lidar are used to perform coordinate conversion to obtain the corresponding 2D image coordinates.

7. The method for recognizing lane lines with large curvature according to claim 6, characterized in that: The RANSAC algorithm is used to extract the road surface 3D point cloud from the 3D point cloud data.

8. The method for recognizing lane lines with large curvature according to claim 7, characterized in that: Lane line extraction from multi-view ground 2D images is performed using deep learning algorithms such as LaneNet, CLRNet, or LaneATT.

9. The method for recognizing a lane line with a large curvature according to claim 8, characterized in that: The DBSCAN point cloud clustering algorithm is used to merge and fuse the multiple 3D lane line coordinate value arrays generated by mapping.

Images