Reconstruction of Human Motions Based on Low-Dimensional Control Signals

Abstract

This thesis addresses the question to what extent it is possible to reconstruct human full-body motions from very sparse control signals.<br /> To this end, we first investigate the use of multi-linear representations of human motions. We show that multi-linear motion models together with knowledge from prerecorded motion capture databases can be used to realize a basic motion reconstruction framework that relies on very sparse inertial sensor input only. However, due to the need for a semantic pre-classification of the motion to be reconstructed and rather restricting database requirements, the described framework is not suitable for a more general motion capture scenario. <br /> We address these issues in a second, more flexible approach, which relies on sparse accelerometer readings only. Specifically, we employ four 3D accelerometers that are attached to the extremities of a human actor to learn a series of local models of human poses at runtime. The main challenge in generating these local models is to find a reliable mapping from the lowdimensional space of accelerations to the high-dimensional space of human poses or motions. We describe a novel online framework that successfully deals with this challenge. In particular, we introduce a novel method for very efficiently retrieving poses and motion segments from a large motion capture database based on a continuous stream of accelerometer readings, as well as a novel prior model that minimizes reconstruction ambiguities while simultaneously accounting for temporal and spatial variations. <br /> Thirdly, we will outline a conceptually very simple yet very effective framework for reconstructing motions based on sparse sets of marker positions. Here, the sparsity of the control signal results from problems that occurred during a motion capture session and is thus unintentional. As a consequence, we do not control the information we can access, which introduces several new challenges. The basic idea of the presented framework is to approximate the original performance by rearranging suitable, time-warped motion subsequences retrieved from a knowledge base containing motion capture data that is known to be similar to the original performance.