Particle filter is a new real time inference algorithm, which is based on Bayesian inference and Monte Carlo method. Because of its unique characteristics such as being flexible, easy to implement, and parallelizable, and being efficient for processing nonlinear problems, particle filter becomes a new and very promising hot topic in applied statistics, signal processing, and artificial intelligence communities. Moreover, it has been applied to many applications such as object tracking and etc. The biggest problem which influences the estimation performance in a particle filter is sample depletion brought by resampling step. This paper focuses on solving this problem from the representation method of particles, and an EM-based Gaussian mixture particle filter is presented. It is demonstrated by computer simulation and visual tracking that the proposed method can reduce the need of sampling numbers and improve the estimation performance of particle filter.