Multi-camera Tracking Based On Particle Filtering

Multi-camera tracking, with more information, can effectively handleocclusions between objects and achieve more accurate and robust tracking.The difficult part of multi-camera tracking is how to extract relevantinformation through different cameras. This paper proposes a multi-cameratracking algorithm based on particle filtering. This algorithm is designed tocope with a calibrated fixed camera network.Contributions of this paper are listed as below. First, the impact oflikelihood function is considered and analyzed based on a numeric experimentproposed by R van der Merwe et al. The result of the experiment shows thatunder some condition the performance of the particle filter can be greatlyimproved with the adjustment of the likelihood function, which is now with noconstraint of the observation model. Second, a knowledge-based backgroundsubtraction method is used in this algorithm in order to tackle the problem oflagging and ghost phenomenon in statistic-based background subtractiontechniques like Gaussian model and Codebook. Third, in the algorithmproposed by this paper, the likelihood function is constructed with a Bayesianmodel, so that the observation of different number of cameras can bemeasured effectively and uniformly.The experiment results display the robust and real-time performance ofthis algorithm, which can accurately handle1to3people in a indoorenvironment, and effectively overcome the problem of severe occlusions. Anintelligent class-recording system based on this algorithm has been put in use.