GD-YOLO: An improved convolutional neural network architecture for real-time detection of smoking and phone use behaviors

Abstract

Chemical safety accidents cause significant socio-economic and environmental hazards, and safety accidents caused by unsafe worker behaviors are preventable. Detection of workers' smoking and phone use behaviors in chemical parks can avoid such safety accidents from the root. Aiming at the issues of reduced accuracy in detecting small objects and low arithmetic capability within industrial equipment. This paper proposes a GD-YOLO network based on YOLOv7 for detecting smoking and phone use behaviors. First, an efficient feature extraction module, D-LAN, is designed to solve the problem of difficult extraction of small object features. Second, a lightweight G-LAN module is proposed to reduce the complexity of the model. Third, WIoU and Mish are introduced to improve the model's performance further, creating a more efficient model. The experimental results show that the mean average precision (mAP) and the number of computational parameters increased by 16.80% and decreased by 21.54%, respectively, compared with YOLOv7. GD-YOLO performs better than other existing mainstream methods in precision and complexity.