Augmented tracking with incomplete observation and probabilistic reasoning

An on-line algorithm for multi-object tracking is presented for monitoring a real-world scene from a single fixed camera. Potential objects are detected with adaptive backgrounds modelled by intensity-plus-chromaticity mixtures of Gaussians to cope with illumination variation. The region-based representations of each object are tracked and predicted using a Kalman filter. A scene model is created to help interpret the occluded or exiting objects. The uncertainty in the domain knowledge is encoded in a Bayesian network for reasoning about object status. Unlike traditional blind tracking during occlusion, the object states are estimated using partial observations whenever available. The observability of each object depends on the predicted measurement of the object, the foreground region measurement, and the scene model. This makes the algorithm more robust in terms of both qualitative and quantitative criteria.