Implementation of Particle Filters for Single Target Tracking Using CUDA

Abstract

In order to implement Sequential Bayesian estimator using Monte carlo simulation and to get rid of limitations of Kalman filter, Particle filtering techniques plays a very crucial role for target tracking applications in state space where Importance sampling approximately distributed by posterior distribution with multimodel feature and robustness to noise. However as the particles becomes very large, the Monte Carlo representation becomes nearly equivalent to analytical description characterization for posterior distributions and has some deficiencies such as high computational cost and low sampling efficiency. Therefore, emerging computing platform, CUDA may be regarded as most appealing platform for such implementaion. Representation provided with set of samples for target distribution of state leads to increase in sampling efficiency so that graphics processing unit (GPUPU) becomes more appealing to use in Particle filter. The modification on mapping architecture are evaluated with qualitative analysis. The proposed design will be 3.5 times faster than direct design.