Traffic sign small target detection model based on improved YOLOv5
Zhipeng Hu, Ying Zhang · 2024
Traffic sign detection, a crucial task in computer vision, has wide-ranging applications in intelligent transportation systems, autonomous driving, and traffic safety. To enhance the performance of detecting traffic sign targets, a model for small target detection is introduced in this study, leveraging an improved version of YOLOv5. By adaptively integrating the Biformer module into YOLOv5, a more flexible content-aware computing distribution can be achieved, thereby enhancing the perception of small targets in traffic signs. NWD loss function is introduced to reduce sensitivity to small position deviations. On the TT100K dataset, the upgraded YOLOv5 model attains an average accuracy of 85%, exhibiting a 3.4% improvement over the original model and showcasing enhanced recognition performance.