Implementation and optimization of Particle Filter tracking algorithm on Multi-DSPs system
Gongyan Li, Bin Li, Zhou Liu, Xiaopeng Chen · 2008
Particle filter is a filter method based on Monte Carlo and recursive Bayesian estimation, which has special advantage in dealing with the state and parameter estimation in the nonlinear and non-Gaussian system. However, high computational complexity and lack of dedicated embedded DSP and ARM hardware for real-time processing have adversely affected its application in real life. In this paper, we present an embedded hardware architecture based on Multi-DSPs (TMS320DM642) for speeding up the basic computational performance, thereby, making Particle Filtering based solutions amenable to real-time constraints. Simultaneously, on this embedded DSP system, we also do some improvement to the particle filter algorithm for realization. First, the number of particles is reduced by fusing mean-shift algorithm after resampling step. Then, RSR (residual systematic resampling) method is mended to reduce the time-consuming division computation and to retain the number of particles same pre-and-post resampling procedure. The performance of the proposed embedded DSP system and optimized algorithm are evaluated qualitatively on real-world video sequences with moving target.