Lane line detection method and system based on sliding window self-attention mechanism

Abstract

The invention relates to a lane line detection method and system based on a sliding window self-attention mechanism, and the method comprises the steps: S1, inputting a foresight traffic image, and carrying out data normalization of the foresight traffic image, and obtaining an input image; S2, inputting the input image into a lane line detection network based on a sliding window self-attention mechanism to obtain different lane line feature maps under different scales; S3, inputting the lane line feature map into a classification network for lane line point classification; carrying out Argmax operation, and taking the position of the point with the maximum probability value as the lane line point of the row; meanwhile, utilizing a probability distribution similarity loss function to constrain the continuous attribute of the lane line; and S4, reflecting the positions of the lane line points to the input image, and outputting the coordinates of the lane line point set. According to the invention, a sliding window-based self-attention mechanism module is provided, the calculation complexity is reduced, the continuous attribute of the lane line is constrained by using a probability distribution similarity loss function, and the lane structure reasoning performance in various scenes is improved.

Description

technical field [0001] The invention relates to the field of automatic driving, in particular to a lane line detection method and system based on a sliding window self-attention mechanism. Background technique [0002] Lane is an important basis for restricting vehicles to drive on the road. Accurate, stable, and fast completion of lane line detection is of great significance for the actual use of autonomous vehicles. With the development of deep learning theory and computing equipment, the lane line detection method based on big data and deep learning has gradually become the mainstream method because of its high detection accuracy and excellent robustness. Convolutional Neural Network (CNN) has achieved great success in computer vision with its powerful feature extraction capabilities. Typical convolutional neural networks, such as VGG, GoogLeNet, and ResNet, have excellent performance in basic tasks including image classification, segmentation, and target detection in co...

AI Technical Summary

Problems solved by technology

[0005] However, the theoretical perceptual domain of the convolutional neural network increases linearly with the number of network layers, and the growth rate is slow. Among them, the effective perceptual domain exists in the form of Gaussian distribution. The perceptual domain has great limitations, and it is difficult to capture long-range dependencies.

This limitation of convolutional neural network leads to the obvious performance degradation of this type of lane line detection method based on convolutional neural network in some difficult scenes.

How to extract the remote dependencies before the lane lines in the image, and then solve the degradation problem in difficult scenes, has become the main bottleneck restricting the development of lane line detection.

Images