A Hybrid Network of CNN and Transformer for 3D Lane Detection

3D lane detection is a critical and challenging task in autonomous driving, yet accurately capturing the spatial structure and precise location of lane markings remains challenging in complex traffic environments. Existing methods typically rely on a single network architecture of Convolutional Neural Networks(CNNs) or Transformer for lane detection, CNNs captures local details but lacks long-range context, while Transformers extract global semantics at the expense of fine local detail, often resulting in imprecise localization. To overcome these limitations, we propose a 3D lane detection framework that integrates both CNNs and Transformer architectures to simultaneously harness local and global information. Specifically, we introduce a view transformation module based on Spatial Cross-Attention to optimize bird’s-eye view (BEV) feature representation. Then we designed a Dual-Branch network, where the CNN branch extracts fine textures and edges, and the Transformer branch captures global semantic relationships. Finally, a feature fusion module (FFM) combining spatial attention and channel attention mechanisms dynamically merges these features to enhance detection performance. Experiments on the OpenLane and Apollo datasets show that our approach significantly outperforms current advanced methods under various complex road conditions, demonstrating its effectiveness and generalizability.