Anomaly detection in videos: A dynamical systems approach

We demonstrate a dynamical system framework based on motion patterns for detecting anomalous individual and group behavior in complex videos. We first describe a framework based on trajectory modeling, in which coarse statistical models are used to capture global motion patterns, and are employed in change detection to identify anomalous behavior at the object level. Our multi-target tracking framework combines geometric active contours with particle filtering to effectively deal with occlusions and clutter in the environment. In crowded scenes, however, such object level representation can become extremely unreliable: to deal with this we instead use of low-level motion features (e.g., optical flow) to capture group behavior. To keep the problem tractable, we utilize a subspace system identification method based on the Hankel matrix to extract relevant low order dynamics of these noisy features. The spectral properties of the Hankel matrix encode useful information about the dynamics, and can detect anomalous group behavior. In order to efficiently compute these spectral properties, we employ a randomized algorithm for singular value decomposition. Both approaches are demonstrated to robustly detect anomalous behavior in realistic indoor and outdoor videos.