SmartAnchor3DLane: Towards monocular 3D lane detection with anchor proposal

Lane detection is an essential technology in intelligent driving systems. Anchor3DLane improves the performance of lane detection by directly sampling the front-view feature map. Still, the most advanced 3D lane detection network relies on fixed pitch and yaw angles to generate lane anchor proposals, making anchor proposal calculation a bottleneck. The span and impact of the yaw angle are large. In this work, we design a yaw proposal network (YPN) and adopt a feature information-sharing mechanism to share feature maps with the network’s backbone, thereby achieving a yaw angle proposal with almost no additional cost. YPN is a simple and lightweight neural network that simultaneously predicts a specific range of yaw angles for each image. YPN is trained end-to-end to generate high-quality yaw proposals, and Anchor3DLane uses these region proposals for detection. Combining the two networks through the feature-sharing mechanism allows the network to recognize the connection between different tasks, thus enhancing the global nature of deep learning tasks. For the popular large-scale real-world lane line dataset OpenLane, 17 proposals are generated for each image. The experimental results show that the proposed method has improved the accuracy compared with the original model, and the average error of the regression of various lane line coordinate values has also been significantly and comprehensively reduced.