Retrieval Of Human Motion Capture Data

The technology of human motion capture takes its advancements along with the development of modern sensor. It has been applied in the areas of vivid animation, health care, clinic diagnosis, motion analysis, robot control, real-scene game, even family entertainment. After the development for more than three decades, the accumulated database of human motion capture data provides advanced support for study and reuse of the human motion capture data. But the human motion capture data is a kind of high-dimensional data and comprises large amount of information, so how to retrieve desired motion sequence efficiently and quickly out of a large human motion capture database remains an unsolved problem in the area of human motion capture.There are two kinds of retrieval: semantic based retrieval and motion based retrieval. Semantic based retrieval means that the retrieval system takes the words specified by user describing the semantic feature of the motion as input, and retrieves the most semanticly relevant motion sequences. Motion based retrieval means that the retrieval system takes a motion sample specified by user as input and retrieves the most numeric similar motion sequence.This paper solves the semantic based retrieval problem by conditional random field (CRF). There are two procedures of the conditional random field: the training procedure and the inference procedure. High level semantic features are needed to be marked out for the training data. K-means is involved into the training procedure to generate the feature functions. The trained CRF model could be used to label new unknown sequences with high level semantic features which are very useful for the indexing procedure afterward. The experiment result shows that this method could label the semantic features out correctly and comprehensively when parameters of the model are chosen appropriately.To solve the motion based retrieval problem, this paper presents an index structure called symbol graph. First, this method transforms the human motion capture data of M dimensions into a K dimensional symbol sequence through equally spatial division. Then hierarchical symbol transformation graph is constructed by the decent order of entropy at each distinct dimension. So the retrieval problem transforms into the intersection of nodes along a path in the symbol graph. Experiment result shows that symbol graph is an efficient index structure.