Modeling Behavior Trends and Detecting Abnormal Events using Seasonal Kalman Filters
J.W. Davis, Mark Keck · 2005
We present a seasonal state-space model using Kalman recursions to learn and predict structured behavior patterns. The model is employed to detect events using the learned expectations of typical scene activity. Abnormal events are detected when a new observation exceeds the confidence range for the predicted behavior. We demonstrate the approach for modeling (over multiple days) the number of pedestrians in a scene, door access card-reader activity, and the departure rate of vehicles from a parking garage. We then use the model to detect abnormal events in each of these domains. The proposed framework provides a single long-term model by exploiting the natural trends in daily human activity. 1