State extraction of probability hypothesis density filter based on Dirichlet distribution

The Dirichlet distribution is applied to extract the multi-target states in the sequential Monte Carlo implementation of probability hypothesis density filter in this paper. The character of Dirichlet distribution with negative exponent parameters, on which this paper is mainly concentrated, is that it makes the components unstable and competitive with each other so that components with less evidence will be discarded during the iteration process. Maximum likelihood criterion is adopted for the proposed algorithm, while it is implemented by the expectation maximum algorithm. In order to reduce the time cost, the k-dimensional tree is applied to initialize the components of Dirichlet distribution. Simulation results prove that the proposed state extraction algorithm is superior to k-means algorithm and expectation maximum implementation of Gaussian mixture model.