Real-Time Multi-Person Smoking Event Detection
Waynebert Jan D. Cabanto, Aira Danielle B. Jocson, Renzel Laurence T. Lateo, Joel C. De Goma · 2019
Smoking is a widespread problem around the globe and having an algorithm that can automatically detect smoking events is crucial in helping to public areas free of dangerous cigarette smoke. This paper presents a method of detecting smoking events through multiple human tracking, action recognition and smoke detection. In the proposed method, human objects are first detected using Histogram of Oriented Gradients (HOG). Then, the detected human objects are tracked using centroid tracking. The smoking event detection requires two factors. The first one is whether the detected human object is displaying actions that would suggest the act of smoking using CNN from Keras library and the other factor is whether there is cigarette smoke in the frame using SVM to identify the candidate areas and color features and pixel values as thresholds. The model obtained an f-score of 72.55% having one person in the frame and 64.37% having 2 or more.