Motion feature filtering for event detection in crowded scenes

We describe a spatio-temporal feature filtering approach that is appropriate for detecting video events in public scenes containing from many to few people. This non-discrete tracking - or pattern flow analysis - is distinguished by the fact that the usual video processing step of object segmentation is omitted; instead motion features alone are used to detect, follow, and separate activity. Motion features include location, scale, score (magnitude), direction, and velocity. The method entails gradient-based motion detection and multiscale motion feature calculation to obtain a scene activity vector. We focus on obtaining these motion features and filtering them to obtain information on activity, with the end-goal being event detection, classification, and anomaly detection. Examples of information extraction we show in this paper include: distinguishing anomalous from trend activity via shape of the activity profile over time, detecting event onset and direction of people flow using direction (and feature confidence) values, and measuring the periodicity of similar activity from magnitude values over time. We demonstrate utility of the approach on 3 video datasets: hallway, emergency event, and subway platform.