Structure-guided feature fusion and geometric angle supervision for robust lane detection

Abstract

Lane detection is a critical task for autonomous driving and advanced driver assistance systems. Existing methods often suffer from insufficient fusion of global structure and local details, as well as underutilization of geometric priors such as vanishing point convergence. This work presents a structure-aware feature fusion attention module to enhance feature pyramid networks, enabling joint modeling of global context and local boundaries for slender lane lines. Meanwhile, an inclination angle loss is designed to embed directional geometric constraints into model training. Experiments on TuSimple and CULane datasets show competitive performance and state-of-the-art results on TuSimple, demonstrating the effectiveness of the proposed components in improving detection accuracy and robustness in complex scenes. The code of our model are available at https://github.com/wlp1213/SINet.