A real-time lane detection network using two-directional separation attention

Real-time network by adopting attention mechanism is helpful for enhancing lane detection capability of autonomous vehicles. This paper proposes a real-time lane detection network (TSA-LNet) that incorporates a lightweight network (LNet) and a two-directional separation attention (TSA) to enhance the lane detection capability of autonomous vehicles. By adopting the attention mechanism, the real-time performance and detection accuracy are significantly improved. Specifically, LNet employs symmetry layer to drastically reduce the number of parameters and the network's running time. TSA infers the attention map along two separate directions, transverse and longitudinal, and performs adaptive feature refinement by multiplying the attention map with the input feature map. TSA can be integrated into LNet to capture the local textural and global contextual information of lanes without increasing the processing time. Results on popular benchmarks demonstrate that TSA-LNet achieves outstanding detection accuracy and faster speed (6.99 ms per image). Additionally, TSA-LNet exhibits excellent robustness in real-world scenarios.