Modified particle implementation of the PHD filter for multi-target tracking
Wei Yang, Yaowen Fu, Jiaqin Zhou, Hongqiang Wang, Xiang Li · 2012
The particle probability hypothesis density (PHD) filter is a popular method for multi-target tracking problem in a Bayesian framework. However, the required resample procedure within this filter is traditionally performed on the whole particle set, which results in the loss of low weighted particles that represent some temporarily miss-detected targets. Another limitation is related to the characteristic of identity free within the particle PHD filter. This paper proposes a modified particle implementation of the PHD filter for multi-target tracking. The proposed method adds extra particle label to create target identities and the resample procedures are performed on the particle set with the same identity label. An identity label management procedure is devised, which sorts tracks into three disjoint subclasses, i.e., tentative, confirmed and terminated tracks. A modified adaptive target birth intensity is adopted to efficiently capture the birth targets that may appear anywhere in the state space. Simulation results demonstrate that the proposed implementation can provide more reliable state estimations and keep record of track identities.