Recursive Bayesian inference and learning for target tracking with unknown maneuvers

This paper addresses the problem of target tracking under completely unknown maneuvering behavior to cope with the complicated target maneuvers. The actual but unknown motion model is formulated as a linear stochastic model with unknown transition matrix and modeling error covariance. Considering the diversity and time-varying of maneuvers, both the transition matrix and covariance are random. Accordingly, the tracker design amounts to recursively updating the joint posterior of target state and unknown parameters (i.e., the transition matrix and covariance). Since it is difficult to obtain the analytical joint posterior, a Bayesian inference and learning scheme is proposed in the sequential Monte Carlo (SMC) framework to alternatively update the state posterior and the unknown parameters posterior. Particularly, it is revealed that the state prediction density used in the SMC obeys a Student's t-distribution, thus ensuring the effective sampling of particle trajectories. Finally, a maneuvering benchmark shows that the proposed tracker can significantly reduce the peak tracking error, which is one of the most important performance indices for preventing tracking failure.