Real-Time Detection of Multi-scale Traffic Signs Based on Decoupled Heads

The efficacy of conventional traffic sign detection is vulnerable to various factors, including diverse lighting, unfavorable weather, and intricate backgrounds. Furthermore, the feature pyramids in deep learning methodologies may destroy the inherent coherence of features among traffic signs, consequently leading to low performance of small object detection. To address these issues, this paper proposes a new model for traffic sign detection, namely YOLOv5-DTR. The model is based onYOLOv5s, enhancing its capacity to detect small targets by constructing additional prediction heads and integrating receptive field enhancement modules. These facilitate the extraction of featuremaps that contain various scales, thereby compensate the potential loss of information. Meanwhile, classification and localization are separated by decoupled heads to alleviate the conflict between the two tasks. In addition, we fuse the triple-attention mechanism with the network, enabling the network to place greater emphasis on the target region. The experimental results on TT100K show that the mAP of the proposed model can achieve 85.9%. Besides, experiments on CCTSDB2021 demonstrate the generalization of this model.