Bingham Distribution-Based Linear Filter for Online Pose Estimation


Authors: Arun Srivatsan Rangaprasad, Mengyun Xu, Nicolas Zevallos-Roberts, Howie Choset

Pose estimation is central to several robotics applicationssuch as registration, hand-eye calibration, SLAM,etc. Online pose estimation methods typically use Gaussiandistributions to describe the uncertainty in the pose parameters.Such a description can be inadequate when using parameterssuch as unit-quaternions that are not unimodally distributed. ABingham distribution can effectively model the uncertainty inunit-quaternions, as it has antipodal symmetry and is definedon a unit-hypersphere. A combination of Gaussian and Binghamdistributions is used to develop a linear filter that accuratelyestimates the distribution of the pose parameters, in their truespace. To the best of our knowledge our approach is the firstimplementation to use a Bingham distribution for 6 DoF poseestimation. Experiments assert that this approach is robust toinitial estimation errors as well as sensor noise. Compared tostate of the art methods, our approach takes fewer iterationsto converge onto the correct pose estimate. The efficacy of theformulation is illustrated with a number of simulated exampleson standard datasets as well as real-world experiments.