Multi-camera Occlusion and Sudden-Appearance-Change Detection Using Hidden Markovian Chains

Abstract

In this paper, a new object tracking algorithm using multiple cameras for surveillance applications is proposed. The proposed algorithm is for detecting sudden-appearance-changes and occlusions. We use a hidden Markovian statistical model, where the random events of sudden-appearance-changes and occlusions are the hidden variables. The tracking algorithm uses both a discriminative model and a generative model for the being-tracked object. The prediction errors in the generative model are used as the observed random variables in the hidden Markovian model. We assume that the prediction errors are exponentially distributed, when no sudden-appearance-changes and occlusion occurs. And the prediction errors are assumed uniformly distributed, when such random events occur. Almost all state-of-the-art discriminative model based object tracking algorithms need to update the discriminative models on-line and thus suffer a so called drifting problem. We show in this paper that the obtained sudden-appearance-changes and occlusion estimations can be used to alleviate such drifting problems. Finally, we show some experimental results that our algorithm detects the sudden-appearance changes and occlusions reliably and can be used for alleviating the drifting problems.

This paper was originally submitted to Xinova LLC as a response to a Request for Invention (RFI) on new event monitoring methods.