Indoor Smoking Behavior Detection Based on YOLOv3-tiny

Abstract

Cigarette smoking is not only harmful to the health of smokers and people around them, but also one of the hidden dangers of indoor fire. Aiming at the problem of heavy workload of manual supervision and low precision of traditional smoke alarm, this paper proposes an improved algorithm based on YOLOv3-tiny deep learning network for indoor smoking behavior detection. After pre-processing data such as sample labeling and partitions data set on sample images, using k-means clustering algorithm to get the size of bounding box priors. In view of the small target of cigarettes and the insignificant difference between smokers and nonsmokers, a small target detection layer added based on the traditional YOLOv3-tiny network. The experimental results show that the Improved YOLOv3-tiny algorithm has the advantages of high accuracy and small model, which can effectively meet the practical application requirements and provide a new way for assisting indoor supervision.