EM-based extended object tracking without a priori extension evolution model

Abstract This paper proposes a novel extended object tracking (EOT) problem, in which the three quantities of object extension (OE), including semi-axis lengths and orientation, are depicted more realistic, rather than assuming prior evolution models as in the traditional methods. By the fact that the semi-axis lengths are intrinsic parameters of object size and the orientation is a state-dependent parameter, OE quantities are treated as parameters to be identified. Such consideration is not only more consistent with the OE practical meaning, but also reduces the uncertainty induced by prior parameter settings. The resultant EOT problem brings out the new challenge: deep coupling between kinematic state and OE, which is difficult to directly derive an Bayesian solution. In the expectation maximization framework, an optimization scheme is developed for joint OE identification and kinematic state estimation. Simulation results show the superiority of the proposed method compared with the state-of-the-art ones.