Research on smoking detection based on deep learning

Abstract

Smoking detection in public places is an important means to protect people's health and safety of life and property, but small target detection will have problems such as low detection accuracy and missed detection. In response to this problem, from the perspective of single-stage detection, a smoking detection model suitable for real-time monitoring is proposed. Based on the custom attention mechanism module and the improved residual network, design a backbone network that reuses underlying features while fusing features at different stages to enhance the ability to extract small target features and improve the accuracy of target detection; the feature fusion part uses FPN The structure and PAN structure are integrated, and a lightweight Neck layer network structure is designed to retain the semantic information and location information of the target feature; the DIOU_nms algorithm is selected to improve the missed detection problem. The self-made smoking data set was detected, the average accuracy rate (mAP) reached 86.32%, and the detection speed reached 55f/s. The detection model in this paper improves the accuracy of smoking detection, improves the supervision effect of smoke-free areas to a certain extent, and provides help for eliminating fire hazards.