Abstract:
Smoking cigarette recognition in a real-time video surveillance system is still an open challenge for computer vision and pattern recognition communities because of its shape, motion, its dependency on lighting conditions, background disturbance, and colors of the scene. Smoking cigarette detection by visual clues can allow fast and reactive alarm systems as compared to smoke detection sensors. A significant number of modern films depict some form of tobacco use, with potentially detrimental effects on the behavior of young people. The proposed project is intended to provide similar services, especially for the university environment. With the development of intelligent real-time video surveillance systems, a relatively newer technology of computer vision has attracted considerable attention. The basic idea is to use image frames of a scene captured by a CCTV camera or any camera to determine if a person is smoking or not. Various image processing and machine learning algorithms are applied to process the captured images frames to determine the smoking cigarette within the visual field of the cameras. Distinguishing features like color, motion, and distance are used to detect smoking gestures and determine whether a person is smoking or not. The system’s results were mixed, partially due to the variety and complexity of the footage used in testing. The system works well in close range and daylight, but accuracy is dropped as the person is far away or if light intensity drops. There is scope for improvement, as a possible avenue of future work. The system can be refined by applying different algorithms for detecting smoke instances.
