Enhancing Anchor-Based Lane Detection with Auxiliary Semantic Segmentation Supervision

Accurate and real-time lane detection is essential for autonomous driving systems. While anchor-based methods such as LaneATT have demonstrated strong performance in both speed and accuracy, they often lack spatial understanding that could be provided by dense semantic segmentation. Conversely, segmentation-based approaches struggle to capture instance-level lane structures. Notably, enhancing vectorized lane detection with auxiliary segmentation supervision has not been widely addressed. Therefore, in this work, we propose a dual-task extension to LaneATT by integrating an auxiliary segmentation head. Our architecture jointly learns vectorized lane representations and pixel-wise lane masks using a shared backbone. It is a hybrid vectorized lane detection model that incorporates auxiliary semantic segmentation to guide spatial reasoning in challenging road scenarios. Experiments on the TuSimple dataset demonstrate that our method improves lane detection accuracy and robustness while preserving real-time performance. Ablation studies further validate the effectiveness of different backbones, segmentation heads, and loss weights. Our proposed approach maintains real-time performance while offering improved spatial consistency.