Minimizing bearing bias in tracking by de-coupled rotation and translation estimates

The problem of tracking Euclidean motion is formulated as a sequential learning of rotations and translations. For tracking modalities such as radar and sonar, this approach avoids a fundamental mismatch that arises with standard trackers that model motion dynamics in Cartesian coordinates but track based on measurements whose noise is best modeled in polar coordinates. By considering motion in terms of rotations and translations and using group-theoretic estimation, the proposed tracker enjoys the advantage of unbiased averaging on the rotation group, in accordance with the geometry of the measurements. We demonstrate the proposed method with illustrative preliminary experiments. The stability and convergence of the proposed algorithm is established, extending known convergence results for online learning of rotations.