Nighttime Lane Detection Based on Retinex Theory and Self-Attention Distillation

In the night scene, the lane marking image is overall dark and the illumination is uneven. These problems lead to low detection accuracy, high missed detection rate, and false detection rate. Aiming at these problems, this paper first uses the unsupervised learning network based on Retinex theory to enhance the night image and improve the contrast of the image, and then introduces self-attention distillation into the efficient residual factorized net (ERFNet) coding layer to make the high-level attention try to train the low-level attention map, to improve the accuracy of lane recognition. At the end of this paper, the method is verified on the constructed nightly Tusimple test dataset. The experimental results show that the lane detection model trained by this method is suitable for night environments and is robust in complex scenes.