Lane Detection in Complex Scenes Based on Deep Snake

Aiming at the challenge of lane detection in complex scenes, a lane detection algorithm based on Anchor, LaneSnakeNet, is proposed. This algorithm generates initial lane anchors through clustering and uses ResNet as the backbone to refine lane prediction across different layers on the output feature maps. A detection module LaneSnake based on self-attention mechanism and global feature fusion has been designed, which integrates global information into lane features and is beneficial for regression in covered lanes. In addition, we design a lane shape similarity loss based on prior information of the lane, focusing on the overall constraints of the lane. Finally, experiments were conducted on the Tusimple dataset, and the results show that the proposed algorithm performs well in lane detection in complex scenes, achieving a speed of 135FPS. Compared with the most advanced Anchor-based methods, it maintains competitiveness and faster speed.