Research on Traffic Marker Detection in Complex Road Scenarios

The rapid development of smart vehicles and autonomous driving technologies has driven the adoption of deep learning-based object detection algorithms for traffic sign recognition in driving assistance systems. However, existing methods struggle to achieve satisfactory performance in terms of detection accuracy for small-scale and long-distance traffic signs. This paper proposes an enhanced traffic sign detection framework based on Faster R-CNN with Feature Pyramid Network (FPN), integrating four improvements to address recognition challenge in complex environments. The framework introduces an Adaptive Intra-scale Feature Interaction (AIFI) module that reconstructs the backbone network through multi-branch feature refinement, enabling dynamic adjustment of receptive fields for optimized local feature extraction. Building upon the enhanced feature representation, we develop a Cross-scale Contextual Fusion Module (CCFM) to improve the Feature Pyramid Network (FPN), effectively aggregating multilevel semantic information. The detection head architecture is further enhanced through dual modifications: RoI Align implementation maintains precise spatial alignment during region-of-interest feature pooling, and Focal Loss function is used to address class imbalance by adaptively weighting hard samples. Comprehensive experiments conducted on benchmark datasets demonstrate the framework's superior performance, achieving a 6.5% increase in mAP50 over conventional baseline model.