YHnet: You Have Two Paths for High-performance 3d Lane Detection

Lane detection is a crucial component of autonomous driving perception systems, directly impacting vehicle navigation and safety. However, in complex scenarios, deep neural networks alone often struggle to achieve accurate and robust detection. To address this challenge, we propose YHNet, a novel dual-branch architecture for monocular 3D lane detection. YHNet integrates deep learning with traditional image processing to enhance detection performance. It features a dual-branch feature extraction mechanism comprising a front-view (FV) feature branch and an edge-enhanced supplementary branch. The FV branch employs a convolutional neural network to extract deep semantic features of the road, while the supplementary branch leverages traditional edge detection techniques to capture road boundaries. These edge features are further processed through convolution and pooling operations to generate complementary representations, improving the FV branch's geometric awareness. To optimize feature utilization, we incorporate a specialized convolutional module that reduces redundancy and enhances feature quality. Additionally, an attention mechanism explicitly models spatial dependencies, improving the positional sensitivity of lane representations. Finally, a view transformation module converts the enhanced FV features into a bird's-eye view (BEV) representation. Extensive experiments on OpenLane, the largest real-world 3D lane detection dataset, demonstrate that YHNet achieves excellent performance in monocular 3D lane detection. The code will be publicly available at https://github.com/idrivingman/YHnet.