SFLR: An Efficient LiDAR-based Lane Detection Network with Small Feature Map and Large Receptive Field

Lane detection plays a crucial role in autonomous driving. Compared to camera-based schemes, LiDAR-based lane detection methods show more robustness to changeable environments. However, LiDAR-based ones have high computational complexity, which makes them difficult to meet real-time requirements. Motivated by that, this paper proposes a LiDAR-based lane detection network with small feature map and large receptive field (SFLR). First, an efficient down-sampling module is developed to reduce the feature map size while minimizing information loss. After that, a novel feature extraction module processes the small-sized feature map to extract features, significantly enlarging the model’s receptive field. Comparative and ablative experiments demonstrate that our approach reduces by 75% floating-point operations (FLOPs) compared to state-of-the-art LiDAR-based methods with comparable F1-score. Additionally, the proposed network has $1.5 \times$ and $5.5 \times$ faster inference speed on CPU and GPU, respectively.