A Research On Abnormal Behavior Detection Methods In Supermarket

As an unskilled crime, shoplifting is happening everywhere in the world. A video-based solution is given in this paper to detect shoplifting in a more efficient way. For detecting shoplifting automatically, the research was conducted from two aspects, human activity recognition and abnormal behavior detection. Our main contributions are three-fold.Firstly, we proposed a hands-position based shoplifting detection method. The customer behavior was defined into 4 states, viewing commodities, taking commodities, putting down commodities and hiding commodities. We used hands position to detect the change between four states. A skin and motion based detector was firstly built for hand detecting. After detect the hands, a regional statistical method was defined for hands tracking. The shopping basket was tracked by CamShift algorithm. When we found the hands position moved to the shelf location, we thought that the state changed from viewing commodities to taking commodities. And after the customer took the commodities, if the hands position didn’t moves back or moves to the shopping basket which means the state change to putting down commodities. And if we found that the hands miss in video or moves to fast, we thought that the state had changed to hiding commodities, which means we detected an abnormal behavior. This detection method can be used to detect the shoplifter who hide the commodities into body or carry-on bags.Secondly, we also presented a method that detecting shoplifting using gaze tracking. Estimated the face region by skin model, we detected the eye position by a gray projection and edge features method. Then the position of pupil was determined by horizontally gray projection. The gaze direction was calculated by the position of eye and the pupil. The shoplifting behavior can be detected by recognizing the behavior of strabismus and change of direction of gaze.Lastly, for human activity recognition methods, a pose-motion model was built for detecting the action of hiding commodities in the shop. A SVM classifier was trained for predicted the action of customer. This method can distinguish the normal shopping actions and abnormal shoplifting actions accurately.