Fast lane detection based on sparse prediction

Lane detection plays an important role in driving safety, so it is necessary to explore more practical lane detection methods. Nowadays, with the development of deep learning, mainstream methods regard the lane detection task as a pixel-level image segmentation task. Although the powerful neural network can greatly improve the detection accuracy, it also brings a large computational cost and reduces the running speed. Therefore, achieving high-precision and high-speed lane detection is still challenging. In this paper, we design an efficient and fast sparse prediction method. Specifically, our method describes the lane line by predicting the location of discrete lane points, so the calculation magnitude is the number of discrete points. This method greatly reduces the calculation cost and is not limited by the number of lane lines. The network structure uses an Encoder–Decoder architecture, and the large receptive field obtained by the structure after downsampling can maintain the spatial continuity between lane points, thus ensuring high accuracy. In addition, self-attention distillation is used to improve the learning ability of the network. Compared with other state-of-the-art methods on the TuSimple and CULane datasets, sparse prediction method achieves high accuracy with obvious speed advantage.