Generative Model for Maneuvering Target Tracking

We consider the challenging problem of tracking highly maneuverable targets with unknown dynamics and introduce a new generative maneuvering target model (GMTM) that, for a rigid body target, explicitly estimates not only the kinematics, here considered as effect variables, but also the underlying causative dynamic variables including forces and torques acting on the rigid body target in a Newtonian mechanics framework. We formulate relationships between the dynamic and kinematic state variables in a novel graphical model that naturally facilitates the feedback of physical constraints from the target kinematics to the maneuvering dynamics model in a probabilistic form, thereby achieving improved tracking accuracy and efficiency compared to competing techniques. We develop a sequential Monte Carlo (SMC) inference algorithm that is embedded with Markov chain Monte Carlo (MCMC) steps to generate probabilistic samples amenable to the feedback constraints. The proposed algorithm can estimate both maneuvering dynamics and target kinematics simultaneously. The robustness and efficacy of this approach are illustrated by experimental results obtained from noisy video sequences of both simulated and real maneuvering ground vehicles.