Smoking Behavior Detection Based on Hand Trajectory Tracking and Mouth Saturation Changes

Smoking seriously endanger people's physical and mental health. In some special places, e.g., gas stations and forest areas, smoking may cause serious accidents. In order to detect smoking behavior timely and accurately, we propose an effective and practical detection method by means of video analysis. Two basic approaches are involved in the proposed method: one is smoke detection; the other is smoking action detection. In the first approach, we detect human face by using a special open source of face detection system called SeetaFace, then segment the mouth area and calculate the corresponding grayscale and saturation, finally we can determine if sudden a change occurs around the mouth. In the second approach, there are two basic steps: firstly, we detect the skin color area based on the skin color ellipse model in YCrCb color space, then determine the initial position of the hand by using the location of the skin color area that relative to the face; secondly, track the trajectory of hand movement by using optical flow and then detect whether the hand overlap the mouth in real time. Finally, we combine the results of the preceding two steps in the second approach with the result in the first approach together and then we can determine smoking or non-smoking person. The experimental results show that the proposed method can effectively detect smoking behavior with a small training sample in real time and achieve the detection rate of 95%.