CN-122024211-A - BEVLane-based lightweight lane line detection method and BEVLane-based lightweight lane line detection system

CN122024211ACN 122024211 ACN122024211 ACN 122024211ACN-122024211-A

Abstract

The application relates to the technical field of automatic driving perception, in particular to a BEVLane-based lightweight lane line detection method and a BEVLane-based lightweight lane line detection system. The detection result is fused with the HD Map/LDMap, so that the stability of the Map is maintained, and the flexibility of coping with temporary changes is realized. The application outputs not only the lane line geometry, but also the topological connection relationship among lanes, and provides rich road structure information for the planning module. Aiming at the characteristics of low-speed and frequent parking of the unmanned logistics vehicle, the receptive field and the key point density of the detection network are optimized, and the berth line detection capability of a close-range parking scene is improved.

Inventors

WANG ZHENJIANG
ZHANG JIANGFENG
ZHU WANGWANG
LIU HONGYONG

Assignees

蜂巢智行(上海)技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. A BEVLane-based lightweight lane line detection method is characterized by comprising the following steps: Acquiring a multi-view image of a vehicle, wherein the multi-view image is acquired by a surrounding camera of the vehicle; Preprocessing the multi-view image to obtain a preprocessed image, wherein the preprocessing comprises de-distortion, chromatic aberration correction and pose synchronization; extracting multi-scale features of the preprocessed image, and projecting the multi-scale features to a bird's-eye view space to obtain a bird's-eye view feature map; Inputting the aerial view feature map into a pre-constructed lane line detection network to obtain a key point thermodynamic diagram, a key point offset, an instance segmentation mask and a topological connection relation of adjacent key points; Generating a lane line detection result based on the key point thermodynamic diagram, the key point offset, the instance segmentation mask and the topological connection relation of the adjacent key points; and outputting a lane line topological structure based on the checking result, wherein the lane line priori structure information is from a high-precision map or a lightweight map.
2. The BEVLane-based lightweight lane line detection method according to claim 1, wherein preprocessing the multi-view image to obtain a preprocessed image comprises: Performing de-distortion processing on the multi-view image based on an internal reference matrix and a distortion coefficient of the camera to obtain a de-distorted image; Performing white balance and exposure consistency adjustment on de-distorted images from different cameras to obtain color difference correction images; and acquiring vehicle pose information of the multi-view image at the acquisition time, and synchronizing the vehicle pose information with the color difference correction image based on a time stamp to obtain a preprocessed image.
3. The BEVLane-based lightweight lane line detection method according to claim 1, wherein extracting the multiscale features of the preprocessed image and projecting the multiscale features to a bird's-eye view space to obtain a bird's-eye view feature map comprises: Extracting multi-scale features of the preprocessed image based on a pre-constructed feature extraction backbone network; and projecting the multi-scale features to the BEV space based on an LSS method or a transducer attention mechanism to obtain a bird's eye view feature map.
4. The BEVLane-based lightweight lane line detection method as claimed in claim 1, wherein the lane line detection network is a lightweight network employing an encoder-decoder, and the training method of the lane line detection network comprises: S1, acquiring a multi-view image sample, and labeling the multi-view image sample to obtain a key point thermodynamic diagram label, a key point offset label, an instance segmentation mask label and a topological connection relation label of the multi-view image sample; S2, extracting a multi-scale feature sample of the multi-view image sample, and projecting the multi-scale feature sample to a bird 'S-eye view space to obtain a bird' S-eye view feature sample; S3, inputting the aerial view feature pattern book into a lane line detection network to obtain a predicted key point thermodynamic diagram, a predicted key point offset, a predicted instance segmentation mask and a predicted topological connection relation of adjacent key points; S4, calculating thermodynamic diagram losses of the predicted key point thermodynamic diagram and the key point thermodynamic diagram label, offset losses of the predicted key point offset and the key point offset label, mask segmentation losses of the predicted instance segmentation mask and the instance segmentation mask label and connection relation losses of the predicted topological connection relation and the topological connection relation label based on a pre-constructed loss function; S5, calculating total loss of the thermodynamic diagram loss, the offset loss, the mask segmentation loss and the connection relation loss based on a pre-constructed loss function; s6, based on the back propagation of the total loss, adjusting parameters of the lane line detection network by combining a gradient descent method; s7, repeating the steps S2-S6 until training is completed.
5. The method for lightweight lane line detection based on BEVLane as claimed in claim 4, wherein the mathematical expression of the loss function is: In the formula, Indicating the total loss of the total of the components, Indicating a loss of the thermodynamic diagram, The equilibrium coefficient representing the thermodynamic diagram loss, Indicating a loss of the amount of the offset, The balance coefficient representing the offset loss, Representing the mask segmentation penalty, The balance coefficient representing the mask segmentation loss, Indicating a loss of the connection relation, The balance coefficient representing the loss of the connection relationship, Representing BEV grids There is a predictive probability of a lane line keypoint, In order to be able to take the focus parameter as such, Representing BEV grids The true label that is the key point is located, As a weight parameter for the background class, Representing the total number of real keypoints in the image, Representing the data from the BEV grid The predicted offset of the center to the actual keypoint, The true offset is indicated by the fact that, Representing the number of grids in which the keypoints are located, Representing the L1 smoothing loss function, Representing the variance loss of the image, Indicating a loss of distance and, Representing the total number of lane line instances in the current image, Represent the first The number of keypoints for each instance, Represent the first The embedded center of the individual instances, Represent the first The predictions of the individual keypoints are embedded into vectors, The radius of the class content is represented, The function of Relu is represented as such, And Are all the lane line instance indexes, Represents the inter-class distance threshold value, Representing lane line examples Is arranged in the center of the (c) in the (c), Representing lane line examples Is arranged in the center of the (c) in the (c), Represents the total number of pairs of keypoints, An index representing any two key points, Representing key points Is a real label of the connection relation of (a), Representing key points Is used for predicting the connection relation probability.
6. The BEVLane-based lightweight lane line detection method as claimed in claim 1, wherein generating lane line detection results based on the keypoint thermodynamic diagram, keypoint offset, instance segmentation mask, and topological connection relationship of neighboring keypoints comprises: Extracting peak points of the key point thermodynamic diagram based on maximum pooling, and taking the peak points with confidence degrees larger than a preset confidence degree threshold value as candidate key points; correcting the candidate key points based on the key point offset to obtain corrected key points; Clustering the plurality of correction key points based on the instance segmentation mask to obtain a plurality of key point clusters; sorting each key point in the key point cluster according to a spatial relationship to obtain a key point sequence, screening key point pairs with connection probability larger than a preset probability threshold value in different key point sequences, and constructing lane topology edges of the key point pairs, wherein the connection probability is extracted from the topology connection relationship of adjacent key points; lane line detection results are generated based on one or more lane topology edges.
7. The method of claim 2, wherein the performing consistency check on the lane line detection result based on the lane line prior structure information to obtain a check result, and outputting a lane line topology structure based on the check result comprises: Adjusting the lane line priori structure information based on the vehicle pose information so as to align the lane line detection result with the lane line priori structure information; When no lane line priori structural information exists at the corresponding position of the lane line detection result, outputting the lane line detection result as a planning basis; when the lane line priori structure information exists at the corresponding position of the lane line detection result, calculating the similarity between the lane line detection result and the lane line priori structure information Wherein the similarity is The mathematical expression of (2) is: In the formula, Representing a hausdorff distance calculation function, As the lane curve in the lane line detection result, A lane curve which is the prior structural information of the lane line, Is distance tolerance; At the similarity When the similarity is larger than or equal to a preset similarity threshold value, outputting the lane line detection result as high confidence, wherein the similarity is that And when the similarity threshold value is smaller than the preset similarity threshold value, judging that the map is inconsistent with the current situation, creating a temporary label for the lane line detection result, and uploading the temporary label to a cloud map service platform.
8. BEVLane-based lightweight lane line detection system, characterized by comprising: The acquisition module is used for acquiring a multi-view image of the vehicle, wherein the multi-view image is acquired by a surrounding camera of the vehicle; The preprocessing module is used for preprocessing the multi-view image to obtain a preprocessed image, wherein the preprocessing comprises de-distortion, chromatic aberration correction and pose synchronization; The feature extraction module is used for extracting multi-scale features of the preprocessed image and projecting the multi-scale features to a bird's-eye view space to obtain a bird's-eye view feature map; The preliminary detection module is used for inputting the aerial view feature map into a pre-constructed lane line detection network to obtain a key point thermodynamic diagram, a key point offset, an instance segmentation mask and a topological connection relation of adjacent key points; The post-processing module is used for generating a lane line detection result based on the key point thermodynamic diagram, the key point offset, the instance segmentation mask and the topological connection relation of the adjacent key points; The lane line detection module is used for carrying out consistency verification on the lane line detection result based on the lane line priori structure information to obtain a verification result, and outputting a lane line topological structure based on the verification result, wherein the lane line priori structure information is from a high-precision map or a lightweight map.
9. An electronic device is characterized by comprising a processor and a memory; The memory is configured to store a computer program, and the processor is configured to execute the computer program stored in the memory, to cause the electronic device to perform the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the method according to any one of claims 1 to 7.

Description

BEVLane-based lightweight lane line detection method and BEVLane-based lightweight lane line detection system Technical Field The application relates to the technical field of automatic driving perception, in particular to a BEVLane-based lightweight lane line detection method and system. Background Lane line detection is an important component of environmental awareness in an automatic driving system, providing lane-level positioning information and travel area constraints for vehicles. Traditional lane line detection methods are usually performed in perspective images, lane lines are extracted through image segmentation or key point detection, and then projected into BEV space through IPM (INVERSE PERSPECTIVE MAPPING ) for use by a planning module. The indirect method has the defects that the IPM transformation depends on the assumption of a flat ground, projection distortion can be generated on a fluctuant road surface or a ramp, and the resolution of a far lane line in a perspective image is low, so that the detection accuracy is reduced. In recent years, some studies have begun to explore lane line detection directly in BEV space, such as BEVLane, by projecting image features into BEV space and processing them, error accumulation in IPM transformation is avoided. However, the existing method is mainly designed aiming at the scene of the passenger car, has insufficient suitability for special requirements (such as low-speed accurate berth, narrow road passage and the like) of the unmanned logistics car, and lacks an effective fusion mechanism with an HD Map (high-precision Map) or LDMap (lightweight Map). Disclosure of Invention Accordingly, the present application is directed to a light-weight lane line detection method and system based on BEVLane to solve the problems in the background art. In order to achieve the above purpose, the present application adopts the following technical scheme: the application discloses a BEVLane-based lightweight lane line detection method, which comprises the following steps: Acquiring a multi-view image of a vehicle, wherein the multi-view image is acquired by a surrounding camera of the vehicle; Preprocessing the multi-view image to obtain a preprocessed image, wherein the preprocessing comprises de-distortion, chromatic aberration correction and pose synchronization; extracting multi-scale features of the preprocessed image, and projecting the multi-scale features to a bird's-eye view space to obtain a bird's-eye view feature map; Inputting the aerial view feature map into a pre-constructed lane line detection network to obtain a key point thermodynamic diagram, a key point offset, an instance segmentation mask and a topological connection relation of adjacent key points; Generating a lane line detection result based on the key point thermodynamic diagram, the key point offset, the instance segmentation mask and the topological connection relation of the adjacent key points; and outputting a lane line topological structure based on the checking result, wherein the lane line priori structure information is from a high-precision map or a lightweight map. In an embodiment of the present application, preprocessing the multi-view image to obtain a preprocessed image includes: Performing de-distortion processing on the multi-view image based on an internal reference matrix and a distortion coefficient of the camera to obtain a de-distorted image; Performing white balance and exposure consistency adjustment on de-distorted images from different cameras to obtain color difference correction images; and acquiring vehicle pose information of the multi-view image at the acquisition time, and synchronizing the vehicle pose information with the color difference correction image based on a time stamp to obtain a preprocessed image. In an embodiment of the present application, extracting a multiscale feature of the preprocessed image, and projecting the multiscale feature to a bird's-eye view space to obtain a bird's-eye view feature map, including: Extracting multi-scale features of the preprocessed image based on a pre-constructed feature extraction backbone network; and projecting the multi-scale features to the BEV space based on an LSS method or a transducer attention mechanism to obtain a bird's eye view feature map. In an embodiment of the present application, the lane line detection network is a lightweight network using an encoder-decoder, and the training method of the lane line detection network includes: S1, acquiring a multi-view image sample, and labeling the multi-view image sample to obtain a key point thermodynamic diagram label, a key point offset label, an instance segmentation mask label and a topological connection relation label of the multi-view image sample; S2, extracting a multi-scale feature sample of the multi-view image sample, and projecting the multi-scale feature sample to a bird 'S-eye view space to obtain a bird' S-eye view feature sample; S3, inpu