A lane detection network based on IBN and attention

Abstract

In intelligent transportation system and advanced driving assistant system, lane detection is an indispensable security link. At present, deep learning has been applied to the task of lane detection, and some methods used semantic segmentation to separate lanes from background. This paper presents a modified encoder-decoder network with instance-batch normalization net (IBN-NET) and attention mechanism based on LaneNet structure. In view of the shortcomings of batch normalization (BN) in capturing texture in end-to-end segmentation, we consider further optimizing this part from the idea of image style transfer, which we solve the problem by replacing pixel-wise classification for scene labeling with capturing content images structure. To take advantage of visual and appearance invariance of instance normalization in encoder stage, IBN layers are applied to replace normal BN layers. Secondly, attention mechanism is added to the network, forcing it to focus on lane regions. This structure is very suitable for two-class semantics segmentation task with only lane and background. The experimental results show that the method can improve detection effect.