Smoking Behavior Detection Based on TF-YOLOv5

In this paper, the determination of smoking behavior is considered as a target localization problem by building a neural network, and the advanced YOLOv5 is chosen as the baseline network. For the problem of low accuracy of YOLOv5 in detecting small targets, an improved TF-YOLOv5 is proposed by adding Transformer modules to the feature extraction part and feature fusion part, respectively, so that the network has the prediction potential of self-attention. In addition, we also integrate Coordinate Attention to find attentional regions in dense scenes, further improving the detection accuracy of our proposed TF-YOLOv5 for small targets in dense scenes and enabling us to achieve better results in the detection of smoking behavior. Experiments on our homemade smoking behavior dataset. show that our proposed TF-YOLOv5 improves 1.8% over the baseline model (YOLOv5).