A Non-linear Pedestrian Tracker Using Velocity-Adaptive Particle Filter with Trajectory Analysis

Object tracking, pedestrian tracking in particular, has a wide range of applications such as video surveillance and pedestrian dynamics. Tracking by detection is a common paradigm for this task. As for tracking pedestrian in videos, most of the existing researches take lots of effort on constant and linear models to estimate pedestrian’s location. Pedestrian’s motion in real life is more likely to follow nonlinear, non-gaussian and multi-modal system rules since they can turn, stop and move at any time. We propose a non-linear pedestrian tracker, using tracking by detection paradigm and developing a velocity-adaptive particle filter to estimate the posterior probability density, so as to estimate given pedestrian motion and track them in video sequence. By the way, the pre-trained detector and simple association method are used for updating the observation template. Our method can achieve better accuracy and precision compared to other filtering tracking method and can successfully track designated pedestrian moving in the presence of occlusion, changeable light intensity and complex dynamic background practically.