Abstract: The classical multiple-model probability hypothesis density filter and its extended methods are based on the sequential Monte Carlo method. For the purpose of capturing an interesting target state, they perform particle prediction based on multiple parallel single-target Markov transition models, and distribute a large number of particles throughout the state space in which the targets may appear. As a result, these classical methods show high computational complexity and poor performance for tracking maneuvering targets. This paper proposes an improved multiple-model probability hypothesis density filter for maneuvering target tracking which estimates the current target Markov transition model probability and model parameters by exploiting the last measurements, and performs particle prediction based on model probability. Consequently, most of the sampled particles will be distributed around the region of likely target appearance in the next frame. Numerical simulations demonstrate that the proposed method not only reduces the number of lost targets, but also improves the accuracy of maneuvering target tracking.

Key words: tracking, probability hypothesis density, particle filter, random finite set, interacting multiple model