On Orientation Distribution Representations with Reliable Uncertainties in Inertial Motion Tracking

Inertial sensor-based human movement capturing systems typically estimate the limb or sensor orientation using a relatively high-frequency stream (usually at least 60Hz) of measurements. However, for many motion analysis tasks, only a so-called segment of a motion, consisting of lower frequencies is analysed and classified. This is typically done via learning a task-specific model that identifies the movement segment, e.g. an action or a gesture. Once movement segment identification models are learned, applying them to obtain low-frequency estimation depends on the accuracy of the uncertainty estimates of movement segment models. To connect both, we start with investigating the accuracy of the distributional representations of orientations using Probabilistic Movement Primitives (ProMP) and Variational Inference (VI) based approximate distributions (e.g., Sparse Gaussian Processes (SGPs)). These techniques are commonly employed to represent movement segment distributions and concisely represent uncertainties. Furthermore, we address challenges associated with orientation regression, based on a case study of human movement data.