Enhancing Indoor Smoking Detection through Deep Learning in AI-Enabled Surveillance Systems

In the post-COVID-19 era, our society increasingly relies on AI-enhanced monitoring and video technologies. These offer significant potential for public safety enhancement. This paper delves into a specific application: the detection of indoor smoking-related illicit behavior. Our research focuses on seamlessly integrating real-time data streaming from CCTV systems with advanced machine learning methodologies. We refine existing deep learning networks to enhance the accuracy of smoking behavior detection. To maintain ethical standards, we implemented Institutional Review Board (IRB) safeguards and upheld data privacy with limited dataset retention. We explored three different modeling techniques: smoke detection, hand-cigarette position detection, and modeling of hand positions during cigarette smoking. While the initial smoke detection technique faced challenges due to CCTV footage complexities, the latter techniques showed great potential. Specifically, by modeling body postures during smoking and employing both deep and classical machine learning models, we achieved remarkable results with a balanced accuracy of 94.41%. These results showcase the potential for robust, real-time indoor smoking detection in standardized environments. Future exploration may involve sequential deep learning models to incorporate temporal information from video data, promising enhanced accuracy and contextual understanding of smoking behaviors.