Efficient monitoring of illicit activities: identifying smokers through human activity recognition

Abstract

The use of security surveillance systems is becoming more prevalent, causing them to become an excellent tool for major crime prevention. However, niche crimes such as the use of cigarettes and other forms of smoking devices in non-smoking environments go unnoticed due to the range of motions required to perform smoking being very minimal. The problem is not only smoking but also the littering that follows afterward. Hence, our research implemented several models that can recognize smoking-related aspects. With these models, it would be possible to detect smokers through surveillance systems and prevent such niche crimes without requiring extensive manpower. In this research, we have used customized visual data to train our models, consisting of videos of smoking from different angles, lighting conditions, and population densities. Object detection models such as YOLOv5, YOLOv8, YOLOv9 and YOLOv10 were used to detect objects related to smoking and classify people as smokers, with YOLOv9 achieving the best results at 96.2% Precision and 76.3% Recall scores. Video Classification was also performed using YOLOv8 to recognize the different features in frames that constitutes to a person being a smoker which achieved a Precision of 90.6% and Recall of 94.2%.