Efficient Multi-Lane Detection Based on Large-Kernel Convolution and Location

Lane detection is the critical sensing technology for autonomous driving systems and advanced driving assistance systems. Along with the rapid evolution of deep learning, vision-based lane detection has made tremendous progress. However, it still faces significant challenges in complex scenarios lacking visual clues, such as severe occlusions and extreme lighting conditions. To detect multiple lanes accurately and efficiently in challenging scenarios, we propose a novel multi-lane detection method called Large-kernel Lane Network (LkLaneNet). By fusing factorized convolution and depth-wise convolution, the Efficient Large-kernel Convolution Module (ELCM) is designed to increase the effective receptive field of the network. Thus, more helpful information can be gathered from a larger region for accurate lane detection. In addition, a location-based instance detection approach is proposed to flexibly distinguish different lanes using the lane start locations at the image boundaries, coupled with the row-wise classification formulation for efficient multi-lane detection. The proposed method is evaluated on two popular lane detection benchmarks, CULane and TuSimple. The results show that our method can achieve advanced performance in complex scenarios while maintaining real-time detection efficiency.