Incorporating Self Attention Mechanism into Semantic Segmentation for Lane Detection

Abstract

Lane detection is a challenging task in the field of vision detection. The annotation information of lane is very sparse, and it is faced with the interference of occlusion, illumination and other factors, which seriously affects the capture of lane features by neural network. In this paper, we propose the Self-Attention Lane Segmentation Network (SALSN) which allows attention-driven, long-range dependency modeling for lane detection task. Although traditional convolutional neural networks have demonstrated their powerful performance, their ability to capture global relationships in images has not been fully explored. We introduce a self-attentive module to model the long-range dependencies between lane features. Lanes have strong shape constraints but weak coherence. In SALSN, we utilize a dense feature fusion framework to better capture lane context information and use all element information to generate lane segmentation images. Experimental results show that SALSN is not only effective in learning the remote dependencies of lane features, but also significantly improves the lane detection performance. We have validated our approach on two large-scale lane detection datasets, and our method can achieve more competitive results.