Anchor Query based Transformer for Lane Detection

Lane detection is a significant task with applications such as adaptive cruise control, lane departure warning, and lane-keep assistance in autonomous driving. Convolutional neural networks (CNN) and transformers have been used for lane detection and achieved great performance. However, it is still challenging under complex scenarios, such as crowded or dazzling conditions. In this paper, a novel end-to-end transformer is proposed for lane detection, in which anchor queries (AQ) are used to provide the lane position priors. A coarse-to-fine (C2F) anchor refinement is implemented in the transformer decoder. In the decoder layer, a novel lane embedding is proposed which combines the features along the anchor and the features along the predicted lanes of the previous decoder layer, and the lane embeddings from different layers are fused through a novel cross-layer attention (CLA). We conducted extensive experiments on the CULane and TuSimple datasets. The results demonstrate that our method achieves state-of-the-art (SOTA) results in complex scenes.