Similarity, Retrieval, and Classification of Motion Capture Data

Three-dimensional motion capture data is a digital representation of the complex spatio-temporal structure of human motion. Mocap data is widely used for the synthesis of realistic computer-generated characters in data-driven computer animation and also plays an important role in motion analysis tasks such as activity recognition. Both for efficiency and cost reasons, methods for the reuse of large collections of motion clips are gaining in importance in the field of computer animation. Here, an active field of research is the application of morphing and blending techniques for the creation of new, realistic motions from prerecorded motion clips. This requires the identification and extraction of logically related motions scattered within some data set. Such content-based retrieval of motion capture data, which is a central topic of this thesis, constitutes a difficult problem due to possible spatio-temporal deformations between logically related motions. Recent approaches to motion retrieval apply techniques such as dynamic time warping, which, however, are not applicable to large data sets due to their quadratic space and time complexity. In our approach, we introduce various kinds of relational features describing boolean geometric relations between specified body points and show how these features induce a temporal segmentation of motion capture data streams. By incorporating spatio-temporal invariance into the relational features and induced segments, we are able to adopt indexing methods allowing for flexible and efficient content-based retrieval in large motion capture databases. As a further application of relational motion features, a new method for fully automatic motion classification and retrieval is presented. We introduce the concept of motion templates (MTs), by which the spatio-temporal characteristics of an entire motion class can be learned from training data, yielding an explicit, compact matrix representation. The resulting class MT has a direct, semantic interpretation, and it can be manually edited, mixed, combined with other MTs, extended, and restricted. Furthermore, a class MT exhibits the characteristic as well as the variational aspects of the underlying motion class at a semantically high level. Classification is then performed by comparing a set of precomputed class MTs with unknown motion data and labeling matching portions with the respective motion class label. Here, the crucial point is that the variational (hence uncharacteristic) motion aspects encoded in the class MT are automatically masked out in the comparison, which can be thought of as locally adaptive feature selection