Adaptive sequential Monte Carlo implementation of the PHD filter for multi-target tracking

In recent years, the sequential Monte Carlo (SMC) implementation of the probability hypothesis density (PHD) filter has been applied with great success in multi-target tracking problem. The standard SMC implementation is equivalent to the particle filter, which involves a mass of particles. Generally, there is a positive correlation between the number of particles and the expected number of targets. However, most of the existing SMC methods use a fixed number of particles per target, which is computationally inefficient. In order to overcome the outlined problem, we propose an adaptive SMC implementation of the PHD (ASMC-PHD) filter. This novel implementation modifies the number of particles adaptively at each time epoch. And the mechanism is realized by comparing the Kullback-Leibler divergence (KL-divergence) with a pre-specified threshold. Accordingly, the number of particles for the next recursion is obtained. Besides, this approach is complementary with the existing SMC methods. Simulation results show that the proposed ASMC-PHD filter based on the KL-divergence is superior to the standard SMC implementation in multi-target tracking.