3D dynamic probabilistic pose estimation from data collected using cameras and reflective markers

Abstract

Position and orientation (Pose) estimations of the human body during motion that are derived from data collected using any marker-based camera system have inherent errors related to a combination of measurement noise, soft tissue artifact (STA), and inaccuracies due to incorrect marker placement. Individually, and in combination, these errors reduce the overall accuracy of marker-based Pose estimation. Optimization and multibody dynamics methods have been formulated to reduce these errors. However it has been argued that uncertainty in data, such as that caused by sensor noise, soft tissue deformation, marker movement, or inaccurate marker placement, cannot be directly accounted for using traditional deterministic approaches. We postulate that uncertainty can be more appropriately addressed by casting the Pose estimation problem within the general framework of probabilistic inference. In this chapter, we will introduce Bayes theorem, the basis for probabilistic inference, and give a general example of how a Bayesian approach can take advantage of prior knowledge to improve estimation. We will then formulate Bayes theorem in the context of mitigating uncertain marker motion. Finally, we will apply this approach on some sample data to demonstrate how this method can, in practice, produce substantially better measurement of knee joint motion then the previously established deterministic methods.