Kalman particle PHD filter for multi-target visual tracking

TLDR: The Kalman filter is applied to generate the proposal distribution, which considers the latest observations in the state transition and matches the posterior density well and provides a robust tracking and outperforms other particle PHD filters.

Abstract: We propose a novel filtering algorithm based on the Probability Hypothesis Density (PHD) for multi-target visual tracking. Some previous methods using particle PHD filter for multi-target tracking have showed superiority in computation and achieved good results, however, the proposal distribution and observation model used in the standard particle PHD filter are naive and poor, which degrade the performance of the tracker. In this paper, the Kalman filter is applied to generate the proposal distribution, which considers the latest observations in the state transition and matches the posterior density well. Moreover, we adopt a precise observation model, which takes the dynamic state of the targets into account, as well as the appearance. The simulation results on real-world scenarios show that our method provides a robust tracking and outperforms other particle PHD filters.