Anomalous Event Detection From Surveillance Video

Content-based video analysis serves as the cornerstone for many applications: video understanding or summarization, multimedia information retrieval and data mining, etc. In our research, we aim to automatically detect anomalous events from surveillance videos (such as video monitoring traffic flow or pedestrian congestion in public spaces). An event is an anomaly if its behavior deviates from what one expects. For example, one such anomaly would be a vehicle making left turn from a straight-only traffic lane. If a system can detect such an event, which poses a safety risk, a human operator can be signaled to alleviate the situation.;Conceptually, what constitutes an anomaly varies in different video scenarios and is difficult to be defined in a general case. Our first solution is based on an unsupervised learning approach. First, all the video events are represented by trajectories of moving objects. Then they are clustered into several behavior patterns under a probabilistic framework. Those patterns with low frequency of occurrence (few trajectory supports) are identified as anomalous patterns. Therefore, our system can automatically detect anomalous object trajectories without acquiring any domain knowledge for different video scenarios. Our contributions include a novel hierarchical clustering algorithm and branch pruning strategies to reduce the complexity.;The second solution extends our anomalous trajectory detection to an arbitrary time length (e.g., one part of a complete trajectory) and multiple objects (multiple trajectories). It is a hierarchal data mining process. We define video events at three semantic levels considering spatiotemporal context: atomic event (motion of one object at any specific time), sequential event (motion of one object within a time range), and co-occurrence event (co-occurrence of multiple objects at specific time). Frequency-based mining techniques are utilized to automatically discover normal event patterns at each level. Those trajectory(ies) parts different from normal patterns are detected as anomalous. Furthermore, we extend this solution to video scenarios where object trajectories cannot be extracted (e.g., crowd motion analysis). Our contributions in this solution include introduction of different event levels and incorporation of spatiotemporal context into video anomaly detection.