YY-YOLO: Improved YOLOv5 for Object Detection on Traffic Signs
Yibing Kong, Yuhan Li, Xinyuan Zhang, Zhiguo Zhou · 2024
As the demand for intelligent transportation and autonomous driving technology increases, traffic sign detection becomes increasingly important. However, in practical applications, factors such as small-sized traffic signs, low resolution, and environmental interference lead to less accurate traffic sign detection. To address this issue, an improved YY-YOLO model is propose d, building upon the foundation of YOLOv5. Firstly, the Leaf Eraser data augmentation algorithm is introduced, which randomly adds tree leaf overlays to the original dataset to address situations where traffic signs are obscured by tree leaves, enhancing the model's robustness. Secondly, a Coordinate attention(CA) mechanism is added to help the model better locate the exact position of the targets. Finally, in the "neck" of the model, the feature pyramid is expanded to larger sizes to optimize the detection performance for small targets. The model demonstrates significant improvements in precision, mean Average Precision (mAP), and mAP at different IoU thresholds.