Pedestrian Tracking Based on Hidden-Latent Temporal Markov Chain

Abstract

Robust, accurate and efficient pedestrian tracking in surveillance scenes is a critical task in many intelligent visual security systems and robotic vision applications. The usual Markov chain based tracking algorithms suffer from error accumulation problem in which the tracking drifts from the objects as time passes. To minimize the accumulation of tracking errors, in this paper we propose to incorporate the semantic information about each observation in the Markov chain model. We thus obtain pedestrian tracking as a temporal Markov chain with two hidden states, called hidden-latent temporal Markov chain (HL-TMC). The hidden state is used to generate the estimated observations during the Markov chain transition process and the latent state represents the semantic information about each observation. The hidden state and the latent state information are then used to obtain the optimum observation, which is the pedestrian. Use of latent states and the probabilistic latent semantic analysis (pLSA) handles the tracking error accumulation problem and improves the accuracy of tracking. Further, the proposed HL-TMC method can effectively track multiple pedestrians in real time. The performance evaluation on standard benchmarking datasets such as CAVIAR, PETS2006 and AVSS2007 shows that the proposed approach minimizes the accumulation of tracking errors and is able to track multiple pedestrians in most of the surveillance situations.