Target tracking for an unknown and time-varying number of targets via particle filtering

Track initialization and adapting to time-varying number of targets are challenges for particle filter (PF) based tracking methods, because of the difficulty of monitoring changes in the entire state space with a finite number of particles. Conversely, if the state space only has limited supports (i.e. after thresholding) or state space is well localized (i.e. target prior information is known), PFs can provide remarkable estimation performance via concentrating its computational power on the areas of interest. In this paper, we present a PF to track unknown and time-varying number of targets based on the track-before-detect (TBD) framework. This PF adopts the two-layer structure of the PF in [5] (i.e. one layer for track initiation, the other for track maintenance), but with different track initiation and maintenance importance densities. It initiates new tracks and controls the false alarms based on the information provided by a maximum a posterior (MAP) method. By discretizing the state space, this MAP-based method can successfully localize the posterior density (i.e. find out the parts of the state space which are of interest) with reasonable computational cost. With regard to the track maintenance, it efficiently updates the existing tracks using a joint, measurement-directed sampling method and deletes the false tracks based on the likelihood ratio test (LRT). Numerical examples are also used to illustrate the efficacy of the proposed algorithm.