Real-Time Lane Detection by Using Biologically Inspired Attention Mechanism to Learn Contextual Information

Background State-of-the-art lane detection methods have achieved prominent performance in complex scenarios, but many limits have also existed. For example, only a fixed number of lanes can be detected, and the cost of detection time is unaffordable in many cases. Methods Inspired by human vision, attention mechanism makes network learning more concerned features. In this paper, we propose a real-time lane detection method by using attention mechanism. The network proposed consists of three modules: an encoder module that extracts the feature of lanes; the instance feature maps of lanes are predicted by two decoder modules, namely binary decoder and embeddable decoder. In the encoder, we use the biologically inspired attention to extract features, which contain many details of the target area. The correlation between the features obtained from the convolutions and that extracted by the attention is established to learn the contextual information. In the decoder, the contextual information is fused with the features from up-sampling, to compensate for the lost detailed information. Binary decoder classifies all the pixels into lane or background. Embeddable decoder obtains the distinguishable lanes. And then, the outputs of the binary decoder serve as one of the inputs to the embeddable decoder to guiding the generation of exact pixel points on the lanes. Results Comparative experiments on two benchmarks (TuSimple and Caltech lanes datasets) show that the proposed method is independent of lane number and lane pattern. It can handle an indefinite number of lanes and run at 10ms in the TuSimple dataset. Conclusions Experiments verify that our method outperforms a lot of state-of-the-art methods while maintaining a real-time performance.