Sequential Monte Carlo Methods for Tracking Multiple Targets With Deterministic and Stochastic Constraints

Abstract

In multitarget scenarios, kinematic constraints from the interaction of targets with their environment or other targets can restrict target motion. Such motion constraint information could improve tracking performance if effectively used by the tracker. In this paper, we propose three particle filtering methods that incorporate constraint information in their proposal and weighting process; the number of targets is fixed and known in all methods. The reproposed constrained motion proposal (RCOMP) utilizes an accept/reject method to propose particles that meet the constraints. The truncated constraint motion proposal (TCOMP) uses proposal densities truncated to satisfy the constraints. The constraint likelihood independent partitions (CLIP) method simply rejects proposed partitions that do not meet the constraints. We use simulation to evaluate the performance of these three methods for two constrained motion scenarios: a vehicle convoy and soldiers executing a leapfrog motion. Moreover, we demonstrate the utility of constraint information by comparing the proposed algorithms with the independent partition (IP) proposal method that does not use constraint information. The simulation results demonstrate that the root mean square error (RMSE) tracking performance of the RCOMP and the TCOMP methods is much better than the CLIP and IP methods; this is due to their more efficient proposal process.