Optimized Tracking Method Using GPU Based Particle Filter

In this paper, we address the problem of object tracking with abrupt motion. In particular, a parallel particle filter processed on Graphic Processing Unit (GPU) with improved proposal distribution has been proposed to generate stable tracking results in real-time. Using maximum likelihood estimation, the observations are introduced into importance function to obtain accurate estimation of the proposal distribution. It reduces the algorithm demand for particle quantity and makes tracking object under unpredictable motion feasible. We transplant particle filter processing into GPU kernel, keeping all the computation on the video card. Paralleling techniques greatly improve the computational efficiency of our proposed algorithm. The experimental results demonstrate that optimized particle filter performances better than classic method, and GPU parallel scheme achieves a good speedup compared to the corresponding sequential algorithms.