CSA-Lanenet: a contiguous spatial attention lane detection network with vision transformer modules

Lane detection is one of the important technologies for automatic driving to effectively avoid the accidents caused by vehicles deviating from their driving lanes. The lane detection task is challenging due to complex scenes and few features in distorted lane lines. Therefore, collecting useful spatial information of lane lines associated with feature maps becomes an important task for precision lane line detection. Considering the image captured from the front car camera, in this paper, we use deep learning segmentation technologies to precisely extract the lane lines by using contiguous spatial attentions. We first propose the shortened spatial attention module into the lane detection network through the correlated spatial local information and transfer to adjacent features to improve lane detection performance. In addition, due to the successful improvement of vision transformers, we also introduce several simplified poolformers into the proposed lane detection network for further improving lane detection exclusively. Simulation results show that the proposed shortened spatial attention module and the simplified poolformer can achieve effective and accurate lane detection. (The code is available at https://github.com/jong12/CSA-Lanet .)