Detection of Snatch Theft Based on Temporal Differences in Motion Flow Field Orientation Histograms

This paper presents a technique for detecting abnormal pedestrian movement for use in video surveillance to automatically detect snatch theft. The proposed technique is based on analysis of the global motion pattern of pixels prior to and after the moment preceding the event that is considered a necessary condition for snatch theft. This technique is based on the premise that the scenes prior to and after snatching will have different movement patterns. Compared to high-level feature detection algorithms which identify and track behavior of individual persons in a scene, the use of low-level features in this technique results in more robust detection as it can handle images where individual pedestrians cannot be properly segmented due to occlusion and other environmental factors. Based on experiments, we are able to detect snatch crime events at a higher rate than other techniques.