A Detection-Aided Multi-target Tracking Algorithm

This paper addresses the problem of tracking multiple objects in monocular video sequences. This problem is difficult because one needs to identify the targets by the finite measurement data, which may be affected by variations of pose, environment clutters, etc. A particle filter based multi-target tracking framework is presented, which operates with a novel observation model. The proposed target appearance model is constructed by combining the local modules that divided according to the structure of human body. In order to depress the sample depletion during the approximation process, the observation information is used to construct the mixture proposal density by integrating the sampling manner guided by detected human faces areas with stochastic dynamic model to generate the new samples. Then the Markov chain Monte Carlo (MCMC) method was employed to recursively estimate the solution of multi-target data association problem. Experimental results show that the proposed tracker can effectively handle complex environments, irregular target motions and partial occlusions to keep the identities of the targets in real world.