HGLFNet: Hybrid Global Semantic and Local Detail Feature Network for Lane Detection

Lane detection is a crucial component of autonomous driving perception systems. Existing methods still encounter performance limitations in complex situations like occlusions and extreme illumination. To address these challenges, this paper introduces a novel Hybrid Global Semantic and Local Detail Feature Network (HGLFNet), designed to enhance lane detection accuracy and robustness in complex scenarios. Technically, we propose a Global Semantic Feature Extraction Module (GSFEM), integrating a Long-Range Information Capture Unit (LRICU) and a Multi-Scale Aggregation Unit (MSAU). GSFEM innovatively utilizes large-kernel convolutions to efficiently capture global contextual information, significantly improving the model’s perception of complete lane line shapes and overcoming the limited receptive field of conventional methods. Furthermore, we develop a Semantic Feature Fusion Module (SFFM) to effectively bridge the semantic gap between global semantic features and local detailed features. SFFM incorporates a Deformable Feature-based Attention Mechanism (DFAM) to adaptively enhance the representational power of fused features and highlight crucial lane line characteristics. HGLFNet effectively integrates global semantic context with local detailed information, enhancing the network’s ability to detect thin and occluded lane line structures. In quantitative evaluation, HGLFNet is compared with state-of-the-art methods. On the VIL-100 dataset, the F1 score and accuracy are improved by 0.75% and 0.38%, respectively, compared to ADNet. On the Tusimple dataset, the F1 score and accuracy are improved by 0.49% and 0.11%, respectively. On the more challenging CULane dataset, the F1 score is improved by 0.03%. The experimental results comprehensively demonstrate that HGLFNet surpasses the existing state-of-the-art techniques in both accuracy and efficiency, providing a novel and effective solution for lane detection in complex scenarios and showing significant potential in practical applications.