LWNet: Line-Shaped Warp Net for Video-Based Lane Detection

Accurate video-based lane detection remains challenging in autonomous driving due to occlusions, illumination variations, and temporal inconsistencies that disrupt cross-frame feature alignment. Existing methods either aggregate multiple past frames—risking the introduction of outdated or noisy information—or rely solely on the previous frame without considering the inherent geometric continuity of lane markings. To address these limitations, we propose LWNet, a Line-shaped Warp Network that introduces a physically guided mechanism for temporally consistent feature refinement. LWNet employs a Line-shaped Warp (LW) module, which constrains deformable convolution offsets within a slender, continuous receptive field aligned with the elongated geometry of lane lines, effectively preventing sampling drift toward surrounding vehicles or background clutter. Integrated into a single-step cross-frame optimization pipeline, the LW module enables robust structural alignment between consecutive frames without the need for long-term temporal aggregation. Experiments on a challenging real-world video lane detection dataset demonstrate that LWNet achieves state-of-the-art performance, outperforming both image-based and video-based baselines in F1-score, precision, and recall. Ablation studies further verify that enforcing moderate geometric constraints significantly enhances temporal stability and detection accuracy. These results show that LWNet provides an efficient, interpretable, and robust solution for video-based lane detection in complex driving environments.