Multi-lanes detection in complex scenes based on Res2Net

Abstract

In order to solve the problems of existing lane detection algorithms with low robustness and low real-time performance in complex road scenes, we propose a grid classification lane detection algorithm based on Res2Net. Lane detection is one of the most important parts of the autonomous driving perception module. Existing lane detection methods based on deep networks mainly treat lane detection as pixel segmentation, which is difficult to meet the real-time requirement as well as solve the problems of lane occlusion, lack, and detection at night or under the glare. We improve the grid classification by considering the lane feature point detection as a grid classification based on the image global features. Since Res2Net can extract multi-scale information, it is embedded in ResNet50 to obtain Res2Net50 as the backbone. By comparing the detection effects of several common feature extraction networks, we concluded that Res2Net50 could extract lane information under various complex road conditions. Extensive experiments on the CULane and TuSimple lane datasets show that our method achieves a 72.54% F1-Measure score on CULane and 95.73% accuracy on TuSimple, processing a 288*800 image in only 10.7 ms. Finally, the lane feature points are fitted using polynomials. Our method helps to improve the accuracy and speed of lane detection for the Advanced Driving Assistance System (ADAS) in complex road scenes.