Enhanced 3D Lane Detection via Dual-Stage Attention and Multi-Scale Dilated Residual Learning

This study presents an improved two-stage 3D lane detection model to overcome the limitations of traditional 2D methods, which lack depth information and struggle in complex road scenarios. The first stage employs a lane segmentation network with a mobile inverted bottleneck convolution module and multi-head attention to enhance feature extraction, effectively addressing the problems of lane occlusion and disappearance. The second stage uses a dual-channel lane prediction network, integrating explicit inverse perspective mapping and implicit learning to improve adaptability to road variations. The feature encoding module incorporates dilated convolutions, residual connections, and multi-scale fusion to accelerate convergence and enhance lane detail extraction. Meanwhile, the position regression module leverages dynamically sparse attention and depthwise separable convolutions to improve sensitivity to lane details and overall representation. Experiments on the Apollo Synthetic Dataset show a 3.4% F-score improvement over the baseline, validating the method’s effectiveness and robustness in 3D lane detection.