Two-Stage Lane Detection Based on Attention Mechanism and Post-Processing of Prediction

Existing lane detection methods have achieved remarkable performance in complex real-world scenarios, but many of them still have some problems, such as poor anti-interference ability, poor real-time performance, and difficult deployment at the vehicle end. In this work, we combine the characteristics of deep learning and traditional algorithms and propose a two-stage lane detection method: Firstly, the model uses the output of backbone based on line-anchor for feature pooling, and at the same time, an attention mechanism based on line-anchor is applied to extract global features. Then, the pooled features are integrated with the global features to solve the problems such as the occlusion of the lane lines. Then, the fused features are input into the classification branch and regression branch, and then the relevant position information such as the confidence degree and offset of the lane line are output. Finally, according to the characteristics of lane detection, a special predictive post-processing module, Post-processing of Prediction Results (PPR) is designed based on the output lane information. PPR can improve the detection performance without increasing the number of model training parameters. The experimental results show that the recognition time of one frame image by our method is only 0.012s, meanwhile the accuracy rate can reach 95.6%.