Sign-YOLO: A Novel Lightweight Detection Model for Chinese Traffic Sign
Weizhen Song, Shahrel Azmin Suandi
Abstract
Traffic sign recognition plays a crucial role in the intelligent vehicle’s environment perception system. However, due to varying weather conditions, illumination, and complicated backgrounds, recognizing traffic signs becomes very challenging. A novel lightweight detection model based on YOLOv5s, namely Sign-YOLO, is proposed to overcome these challenges. Firstly, the CA (Coordinate Attention) module is incorporated into the backbone network to improve the extraction of key features. Secondly, the improved High-BiFPN is used to enhance YOLOv5s’ neck structure’s capability in fusing multi-scale semantic information. Finally, the improved Better-Ghost Module is employed to reduce the model’s parameters and accelerate the detection speed. We used the CCTSDB2021 dataset to evaluate our model. Compared to YOLOv5s, the proposed Sign-YOLO algorithm in this paper reduces the model parameters by 0.13 M. The precision, recall, F-1 score, and mAP value have improved by 1.02%, 7.01%, 1.84%, and 4.61%, respectively. The FPS value remains around 86 fps. The results show that Sign-YOLO has achieved the optimal balance between accuracy and real-time performance.