A smoking behavior detection method based on the YOLOv5 network

Abstract

Smoking in public places not only brings about some safety hazards, but also does harm to people's lives, property and living environment. A smoking behavior detection model based on deep learning is trained for the concern of environment and safety. First, a vertical rotation data enhancement method is adopted in the preprocessing stage to extend the dataset and increase the objects of detection. Then, the channel attention module is introduced in backbone network to calibrate the feature response. Finally, added a small target detection layer to the YOLOv5 algorithm. This paper analyzes the network structure of the YOLOv5s, and the model is trained and tested by utilizing the YOLOv5s network. Experimental results show that the mAP value of the algorithm is improved by 5.3% over the original algorithm.