CSGNet: A LiDAR-Based Lane Detection Network With Cyclic-Shifting Group Convolution

Lane detection is one of the most critical tasks in autonomous driving. In the past few years, due to the development of deep neural networks, lane detection approaches using onboard sensors like cameras and LiDAR have been proven to be effective ways to improve performance. In contrast to the camera-based scheme, LiDAR-based lane detection exhibits remarkable robustness to varying lighting conditions. This paper proposes a novel cyclic-shifting group convolution (CSGConv) module. Compared with classical group convolution, the proposed CSGConv module can efficiently promote information exchange among different group feature channels and reduce the high computational burden in current LiDAR-based lane detection networks. Then, a cross stage partial CSGConv (CSPCSG) block is designed to enlarge the receptive field and improve detection performance, especially when the lanes are curved or occluded. Subsequently, a LiDAR-based lane detection network, CSGNet, is put forward by integrating the CSPCSG block. Experiments and ablation tests on the dataset illustrate that our strategy achieves the highest 83.9% overall F1-score. Compared with the current state-of-the-art LiDAR-based lane detection method LLDN-RW, our results exhibit 2.5 times faster and 39% reduction of FLOPs, which indicates less resources are required in the online process.