Citation

Li W, Zhang G, Cui L. J. Transp. Eng. A: Systems 2023; 149(5): e04023025.

Copyright

(Copyright © 2023, American Society of Civil Engineers)

DOI

10.1061/JTEPBS.TEENG-7461

PMID

unavailable

Abstract

How to strike a balance between detection speed and recognition accuracy has become a major challenge in real-time object detection. In this research, the YOLOv5 (You Only Look Once version 5) model was lightweight and optimized to improve the detection speed and accuracy of the network. To prune the backbone and neck are to simplify the network structure and reduce the parameters. The lightweight structure of the C3 module was designed and incorporated into the attention mechanism to improve the feature extraction capability of the network. For the public traffic sign dataset, the label assignment strategy and the loss function of YOLOv5 were refined to alleviate the imbalance between positive and negative samples and to better compute the loss, resulting in more stable and efficient training. Compared with other mainstream single-stage models, it achieves a better trade-off between speed and accuracy. With only 0.85 M parameters, 91.9% of mAP (mean average precision) and 360 FPS (frames per second) were achieved, which were 16.26% mAP and 26.67 FPS higher than the conventional YOLOv5n, respectively. The performance of our lightweight model in traffic sign detection far exceeds the most advanced achievements.

Language: en

Keywords

Coordinate attention; Label assignment; Lightweight; Loss function; Traffic sign recognition; You Only Look Once version 5 (YOLOv5)