CN-122024190-A - Lane line detection method and device

CN122024190ACN 122024190 ACN122024190 ACN 122024190ACN-122024190-A

Abstract

The invention relates to the technical field of intelligent driving lane line detection and discloses a lane line detection method and device, wherein the method comprises the steps of obtaining an original road image to be detected; extracting lane semantic features of an original road image to be detected to obtain a feature map, performing transform feature enhancement on the feature map to obtain a reinforced feature map, performing automatic anchor point distribution embedding on the reinforced feature map to generate a query embedding tensor, and predicting to obtain lane line detection information based on the reinforced feature map, the query embedding tensor and a position code obtained by training in advance. The invention innovatively introduces a dynamic updating automatic anchor point allocation embedding mechanism, replaces the traditional method of manually setting anchor points, and can adaptively learn the number and distribution of the lane lines, thereby stably supporting the identification of at least 3 lane lines on one side, and solving the problems that the traditional detection method is difficult to meet the real-time detection requirement and is difficult to realize the identification of multiple lane lines on one side.

Inventors

ZHOU YUNLONG
CUI ZHANPENG
CHEN MO
JIA JIA
HUANG YA
YU XIONGFENG

Assignees

中国人民解放军63921部队

Dates

Publication Date: 20260512
Application Date: 20251215

Claims (10)

1. A lane line detection method, the method comprising: Acquiring an original road image to be detected; extracting lane semantic features of the original road image to be detected to obtain a feature map; performing transform feature enhancement on the feature map to obtain an enhanced feature map; Performing automatic anchor point allocation embedding on the enhanced feature map to generate a query embedding tensor; and predicting to obtain lane line detection information based on the reinforcement feature map, the query embedded tensor and the position codes obtained by pre-training.
2. The lane-line detection method according to claim 1, wherein before extracting the lane semantic features of the original road image to be detected, the method further comprises: and performing image cutting, interference background elimination and zoom resolution on the original road image to be detected to obtain a road image with fixed resolution.
3. The lane line detection method according to claim 1, wherein extracting the lane semantic features of the original road image to be detected to obtain a feature map comprises: and carrying out hierarchical feature extraction and fusion on the original road image to be detected by adopting a residual network to obtain a feature map containing lane semantic features.
4. The lane line detection method according to claim 1, wherein performing a transform feature enhancement on the feature map to obtain an enhanced feature map comprises: and enhancing global context of the lane line semantic features in the feature map by adopting a transducer structure to obtain an enhanced feature map.
5. The lane-line detection method of claim 1 wherein performing automatic anchor point allocation embedding on the reinforcement feature map, generating a query embedding tensor, comprises: extracting the line number of the running and train tracks, the anchor point number of the lane lines and anchor point coordinate information from the reinforced feature map; and generating query embedded tensors by adopting a multi-layer perceptron based on the number of the lines of the running and train tracks, the number of the anchor points of the lane lines and the anchor point coordinate information.
6. The lane-line detection method according to claim 5, wherein the query embedded tensor is expressed as follows: ; Wherein, the The method comprises the steps of embedding tensors for query, wherein MLP is a multi-layer perceptron operator, concat is a stack of forward and backward tensors and column tensors in one dimension; 、 The number of the running and train tracks respectively; 、 the number of row and column anchor points respectively; 、 coordinate information of x and y in the line anchor points of the line and the train respectively.
7. The lane-line detection method according to claim 1, wherein the predicting the lane-line detection information based on the reinforcement feature map, the query embedded tensor, and the pre-trained position code includes: and carrying out fusion prediction through a attention mechanism and a multi-layer perceptron based on the reinforcement feature map, the query embedded tensor and the position code obtained by training in advance to obtain lane line detection information, wherein the lane line detection information comprises lane line positions, lane line types and confidence information.
8. The lane line detection method according to claim 1, wherein the method further comprises: And carrying out post-processing of a model training stage based on the lane line detection information.
9. The lane line detection method according to claim 8, wherein the post-processing of the model training stage based on the lane line detection information includes: Based on the lane line detection information, dynamically updating the query embedded tensor through post-processing and the multi-layer perceptron, and outputting iterative optimization model parameters based on the updated query embedded tensor and the model until the model converges.
10. A lane line detection apparatus, characterized in that the apparatus comprises: The image acquisition module is used for acquiring an original road image to be detected; the feature extraction module is used for extracting lane semantic features of the original road image to be detected to obtain a feature map; the feature enhancement module is used for carrying out transform feature enhancement on the feature map to obtain an enhanced feature map; the query embedding tensor generation module is used for executing automatic anchor point allocation embedding on the enhanced feature map to generate a query embedding tensor; the decoding module is used for predicting and obtaining lane line detection information based on the reinforcement feature map, the query embedded tensor and the position codes obtained through pre-training.

Description

Lane line detection method and device Technical Field The invention relates to the technical field of intelligent driving lane line detection, in particular to a lane line detection method and device. Background Before the carrier rocket and the spacecraft formally enter the space launch field, a transportation test is usually carried out, so that the trafficability of the transportation path of the carrier rocket and the spacecraft is ensured. The traditional transportation test is manually carried out, the test efficiency and the test precision are low, and the ever-increasing space launching requirements can not be met. By integrating a laser radar, a visible light camera, combined navigation equipment and the like, a detection device aiming at ground typical traffic targets and air typical obstacles can be formed to replace a traditional transportation test of manual operation. The structured road lane line is used as boundary constraint and input condition of the detection device, real-time, stable and accurate detection results are required, and meanwhile, at least 3 lane lines are required to be identified on one side based on ultra-wide overrun characteristics of the carrier rocket and the spacecraft and the working mode of the detection device. At present, the main stream lane line detection method mainly comprises a lane line detection method based on semantic segmentation and a lane line detection method based on anchor point regression. The lane line detection method based on semantic segmentation has strong feature extraction capability, but has large algorithm model size and low reasoning speed, and is difficult to meet the real-time detection requirement, and the lane line detection algorithm based on anchor point regression adopts a fixed anchor point allocation strategy, so that anchor points are required to be manually set, and single-side 3 lane lines are difficult to identify. Therefore, a lane line detection method suitable for ultra-wide roads is needed. Disclosure of Invention The invention provides a lane line detection method and a lane line detection device, which are used for solving the problems that the traditional detection method is difficult to meet the real-time detection requirement and is difficult to realize single-side multi-lane line identification. In a first aspect, the present invention provides a lane line detection method, including: Acquiring an original road image to be detected; extracting lane semantic features of an original road image to be detected to obtain a feature map; performing transform feature enhancement on the feature map to obtain an enhanced feature map; Performing automatic anchor point allocation embedding on the enhanced feature map to generate a query embedding tensor; and predicting to obtain lane line detection information based on the reinforcement feature map, the query embedded tensor and the position codes obtained by pre-training. The lane line detection method provided by the invention has the advantages that the strong global context modeling capability of a Transformer is utilized, the receptive field of the feature map to the lane lines is obviously enhanced, the long-distance dependence relationship can be effectively captured, the robustness to complex road conditions and shielding conditions is improved, a dynamically updatable automatic anchor point allocation embedding mechanism is innovatively introduced, the traditional method of manually setting anchor points is replaced, the number and distribution of the lane lines can be adaptively learned, the identification of at least 3 lane lines on one side is stably supported, the path detection requirements of overrun transport means such as carrier rockets are perfectly matched, the processing performance is ensured, meanwhile, the detection accuracy and environmental adaptability are greatly improved, the reliable lane perception capability is provided for intelligent driving and special vehicle automatic driving, and the problems that the real-time detection requirement and the identification of multiple lane lines on one side are difficult to meet in the traditional detection method are solved. In an alternative embodiment, before extracting the lane semantic features of the original road image to be detected, the method further comprises: And performing image cutting, interference background elimination and zoom resolution on the original road image to be detected to obtain a road image with fixed resolution. According to the lane line detection method provided by the invention, the image cutting, the interference background removing and the zoom resolution are carried out on the original road image to be detected, the road image with fixed resolution is obtained, the irrelevant background interference above the lane line is effectively removed, the attention of a model is focused on a road area, the false detection risk is reduced, the stable and standard input is provided f