Assessing Temporal Coherence for Posture Classification with Large Occlusions

In this paper we present a people posture classification approach especially devoted to cope with occlusions. In particular, the approach aims at assessing temporal coherence of visual data over probabilistic models. A mixed predictive and probabilistic tracking is proposed: a probabilistic tracking maintains along time the actual appearance of detected people and evaluates the occlusion probability; an additional tracking with Kalman prediction improves the estimation of the people position inside the room. Probabilistic Projection Maps (PPMs) created with a learning phase are matched against the appearance mask of the track. Finally, an Hidden Markov Model formulation of the posture corrects the frame-by-frame classification uncertainties and makes the system reliable even in presence of occlusions. Results obtained over real indoor sequences are discussed.