End-to-end lane detection with convolution and transformer

In this paper, an end-to-end lane detection method based on the polynomial regression is proposed, combining CNNs and Transformer. Transformer proposes a self-attentive mechanism to model nonlocal interactions to capture global context. Then, an effective Global-Local training strategy is presented to capture a multi-scale feature, which is capable of capturing richer lane information involving structure and context, especially as the lane marking point is remote. And the obtained multi-scale feature map can be fused by utilizing different scale guidance. Finally, the proposed method is validated on the TuSimple benchmark, whose results show the accuracy can achieve 96.33% in models, and 11.1x faster than the popular Line-CNN model in "compute" time.