Multi-target tracking in video by SMC-PHD filter with elimination of other targets and state dependent multi-modal likelihoods

Fast algorithm of SMC-PHD filter for track-before-detection in visual tracking by eliminating other targets from the original image has been proposed. The elimination allows us to approximate the original likelihood term in PHD filter to an simplified one leading to a fast algorithm having computational complexity proportional to the number of particles, not to the number of particles by the number of observations as in conventional PHD filters. Additionally, we propose state dependent multi-modal likelihoods to represent multiple intensities coming from different mode of sensors that depend on, such as size of target, posture of the target, and so on. Experiments to track multi-pedestrians in real street scene in a video captured by a car cabin camera demonstrate performance of proposed method. Comparisons with conventional method, which has no state dependent multi-modal likelihoods idea, have been conducted and we can conclude that the proposed method can provide a proper number of trajectories of pedestrian tracking.