Lane line detection based on the codec structure of the attention mechanism

For self-driving cars and advanced driver assistance systems, lane detection is imperative. On the one hand, numerous current lane line detection algorithms perform dense pixel-by-pixel prediction followed by complex post-processing. On the other hand, as lane lines only account for a small part of the whole image, there are only very subtle and sparse signals, and information is lost during long-distance transmission. Therefore, it is difficult for an ordinary convolutional neural network to resolve challenging scenes, such as severe occlusion, congested roads, and poor lighting conditions. To address these issues, in this study, we propose an encoder-decoder architecture based on an attention mechanism. The encoder module is employed to initially extract the lane line features. We propose a spatial recurrent feature-shift aggregator module to further enrich the lane line features, which transmits information from four directions (up, down, left, and right). In addition, this module contains the spatial attention feature that focuses on useful information for lane line detection and reduces redundant computations. In particular, to reduce the occurrence of incorrect predictions and the need for post-processing, we add channel attention between the encoding and decoding. It processes encoding and decoding to obtain multidimensional attention information, respectively. Our method achieved novel results on two popular lane detection benchmarks (CULane F1-measure 76.2, TuSimple accuracy 96.85%), which can reach 48 frames per second and meet the real-time requirements of autonomous driving.