CCLane: Concise Curve Anchor-Based Lane Detection Model with MLP-Mixer

Lane detection needs to meet the real-time requirements and efficiently utilize both local and global information on the feature map. In this paper, we propose a new lane detection model called CCLane, which uses the pre-set curve anchor method to better utilize the prior information of the lane. Based on the Cross Layer Refinement method for extracting local information at different levels, we propose a way to combine MLP-Mixer and spatial convolution to obtain global information and achieve information transmission between lanes, which flexibly and efficiently integrates local and global information. We also extend the DIoU loss function to lane detection and design the LDIoU loss function. The method is evaluated on two widely used lane detection datasets, and the results show that our method performs well.