Lane Detection Method Based on MCA-UFLD

Lingyun Han, Kun Xu, Wensheng Hu, Zhanwen Liu

Abstract: Lane detection is a key technology in automatic driving systems, and one of the mainstream methods is detection based on segmentation. However, this method requires pixel by pixel segmentation of the image. It cannot demonstrate good real-time and robustness when facing complex scenarios. This paper proposes a lane detection method based on MCA-UFLD network, which defines lane detection as the location selection and classification of predefined row anchors in the row direction. The method greatly improves the detection speed. In this paper, the lightweight network MobileNetV2 is used as the backbone network to reduce the parameters needed by the network and improve the running speed. The coordinate attention (CA) module is introduced into the backbone network, and the location information is embedded to enhance the feature extraction of important lane information and solve the problem of long-distance dependence of lanes. Aiming at the problem that the accuracy of long-distance lane detection is not ideal in some complex scenes, an auxiliary branch of vanishing point detection is added to help identify the position and direction of lanes. In this paper, the proposed network is tested on the lane detection dataset CULane, and the performance is compared with the existing classical lane detection networks. The experimental results show that the F1-measure of the MCA-UFLD network on the CULane dataset reaches 69.36, and the FPS reaches 77.14. Compared with RESA, ERFNet, and UFLD network with ResNet-152 as the backbone, the F1-measure is increased by 0.18, 0.53, and 0.38 percentage points respectively, and the FPS is increased by 23.15, 18.18, and 48.49 respectively.