GraphLane: a spatio-temporal graph learning framework for lane localization

Lane localization is an important component of advanced driver assistance systems and autonomous driving, where accurate estimation of lane geometry is essential for safe vehicle navigation. Nevertheless, the current lane detection algorithms may face difficulties in complex driving situations due to poor visual feature extraction, unstable temporal prediction, and inadequate modeling of structural relations between lanes.

In order to overcome these issues, this paper introduces the unified spatio-temporal graph learning approach that integrates spatial feature extraction, temporal context aggregation, uncertainty-driven graph formation, graph attention reasoning, and geometric optimization. The proposed GraphLane framework dynamically learns graph connections between lane keypoints while incorporating uncertainty estimation to suppress unreliable feature interactions and temporal consistency to improve structural stability across consecutive frames. A geometry-aware optimization strategy is further introduced to encourage smooth and physically consistent lane predictions. The framework was evaluated on the CULane test set using the proposed row-anchor-based lane-existence and localization evaluation.

Experimental results show that the proposed GraphLane framework achieves a lane-existence precision of 0.9769, recall of 0.8990, F1-score of 0.9363, and an average localization error of 19.28 pixels. Qualitative analysis shows that the framework produces smooth and consistent lane predictions under occlusions, shadows, dense traffic, and nighttime conditions.

Other analyses on learned graphs, computational complexity, and ablation studies also show that incorporating temporal context, adaptive graph reasoning, and geometrically informed optimizations is effective for lane localization.