Fusion of nonlinear motion dynamics using Fokker-Planck equation and projection filter

This paper presents a novel approach to the fusion of nonlinear motion dynamics for 2D target tracking applications in video analytics by using Fokker-Planck equation (FPE) and the projection filter. The motion dynamics of the target is first conveniently represented using the corresponding FPE for the state posterior. The motion dynamics is then fused into target tracking process to reliably predict the target position and velocity in 2D target tracking by solving the corresponding FPE of the state posterior using a projection filter. Once the next system measurement is available, the state posterior is then updated using the Bayes' rule in the projection filter framework as well. Experiments using synthetic and real aerial surveillance video data show that the proposed FPE-based target tracker is able to reliably track targets in the present of nonlinear motion dynamics, and that the proposed FPE-based tracker outperforms traditional nonlinear filters in target tracking such as the Kalman filter (including extended Kalman filter and unscented Kalman filter) and the particle filter.