Traffic Sign Recognition Algorithm Based on Improved YOLOv5s
Xun Liu, Xiangkui Jiang, Haochang Hu, Rui Ding, Hong Li, Chunlin Da
Abstract
To resolve the problem that the YOLO(You Only Look Once)v5s algorithm model is difficult to accomplish the traffic sign recognition task in small target detection, this paper proposes a YOLOv5s-MobileNetV2 algorithm, replacing the original backbone network DarkNet-53 of YOLOv5s with MobileNetV2 network for feature extraction, and selecting Adam as the optimizer. TT100K dataset containing information of up to 221 traffic sign categories was selected for training. The testing results show that compared with YOLOv5s algorithm, the algorithm in this paper reduces the parameter amount by 65.6% and the computation amount by 59.1% based on the mAP(mean Average Precision) improvement of 0.129.