Human tracking in video surveillance using particle filter

Automated human tracking is a task that has a wide area of applications and has become more important nowadays. This research proposes to investigate the use of Bayesian inference technique specifically particle filter for tracking human in video surveillance. Kalman filter which has been the de facto technique for real world tracking performs poorly for most of the problems because, the real world applications are often non-linear and non Gaussian. The particle filter on the other hand is a tool for estimating the posterior probability density of state of a dynamic model that includes non-linear and non-Gaussian real world applications. The filter uses random sample to estimate the possible location of the tracked object in the next immediate frame even in the presence of occlusion. In order to initialize the tracking process, humans are first detected using a pretrained human detection model in video. The detector utilize model fusing method which is the combination of histogram of oriented gradient based human detector model and Haar feature based upper body detector to locate position of moving person in video. The technique performed excellently well when evaluated on the publicly available CAVIAR dataset and outperformed the Kalman filter algorithm.