Finding Kinematic Structure in Time Series Volume Data

Abstract

This paper presents a new scheme for acquiring 3D kinematic structure and motion from time series volume data. Our basic strategy is to first represent the shape structure of the target in each frame by Reeb graph which we compute by using geodesic distance of target's surface, and then estimate the kinematic structure of the target which is consistent with these shape structures. Although the shape structures can be very different from frame to frame, we propose to derive a unique kinematic structure by way of clustering some nodes of graph, based on the fact that they are partly coherent to a certain extent of time series. Once we acquire a unique kinematic structure, we fit it to other Reeb graphs in the remaining frames, and describe the motion throughout the entire time series. The only assumption we make is that human body can be approximated by an articulated body with certain numbers of end-points and branches. We demonstrate the efficacy of the proposed scheme through some experiments.