THE UNSCENTED KALMAN PARTICLE PHD FILTER FOR JOINT MULTIPLE TARGET TRACKING AND CLASSIFICATION

摘要

The probability hypothesis density (PHD) is the first orderrnstatistical moment of the multiple target posterior density;rnthe PHD recursion involves multiple integrals that generallyrnhave no closed form solutions. A (Sequential MonternCarlo)SMC implementation of the PHD filter has been proposedrnto tackle the issue of joint estimating the number ofrntargets and their states. However, because the state transitionrndoes not take into account the most recent observation,rnthe particles drawn from prior transition may have very lowrnlikelihood and their contributions to the posterior estimationrnbecome negligible. In this paper, we propose a novel algorithmrnnamed Unscented Kalman Particle PHD filter (UKP-rnPHD), and associate it with Multiple dynamical Modelsrn(MM)method. The algorithm consists of a P-PHD filter thatrnuses an Unscented Kalman filter to generate the importancernproposal distribution; the UKF allows the P-PHD filter to incorporaternthe latest observations into a prior updating routinernand thus, generates proposal distributions that match the truernposterior more closely. Moreover, The MM solves the problemrnof tracking manoeuvring targets. Simulation shows thatrnthe proposed filter outperforms the P-PHD filter.