Robust 3D lane detection in complex traffic scenes using Att-Gen-LaneNet

Robust 3D lane detection is the key to advanced autonomous driving technologies. However, complex traffic scenes such as bad weather and variable terrain are the main factors affecting the robustness of lane detection algorithms. In this paper, a generalized two-stage network called Att-Gen-LaneNet was proposed to achieve robust 3D lane detection in complex traffic scenes. The Efficient Channel Attention (ECA) module and the Convolutional Block Attention Module (CBAM) were combined in this network. In the first stage of the network, we improved the semantic segmentation network ENet and proposed the weighted cross-entropy loss function to solve the problem of ambiguous distant lane segmentation. This method improved Pixel Accuracy to 99.7% and MIoU to 89.5%. In the second stage of the network, we introduced the interpolation loss function to achieve accurate lane fitting. This method outperformed existing detection methods by 6% in F-score and Average Precision on the Apollo Synthetic dataset. The proposed method achieved better overall performance in 3D lane detection and was applicable to broader and more complex traffic scenes.