Extended Object Tracking with Spatial Model Adaptation Using Automotive Radar

Abstract

This paper considers extended object tracking (EOT) using high-resolution automotive radar measurements with online spatial model adaptation. This is motivated by the fact that offline learned spatial models may be over-smoothed due to coarsely labeled training data and can be mismatched to onboard radar sensors due to different specifications. To refine the offline learned spatial representation in an online setting, we first apply the unscented Rauch-Tung-Striebel (RTS) smoother that explicitly accounts for the predicted and filtered states based on the offline learned model (i.e., the B-spline chained ellipses model). The smoothed state estimates are then used to create an online batch of state-decoupled training data that are subsequently utilized by an expectation-maximization algorithm to update the spatial model parameters. Numerical validation with synthetic automotive radar measurements is provided to verify the effectiveness of the proposed online model adaptation scheme.