A Comparison And Evaluation Of Motion Indexing Techniques

Motion capture (mocap) is a way to digitally represent the spatio-temporalstructure of human movement. Human motion is generally captured in long sequences to record the natural and complex aspects of human actions, its sequential transitions, and simultaneous combination. As the amount of mocap data increases, it becomes important to index the data in order to improve the performance of information access. The motion indexing problem involves several challenging issues due to the data being high dimensional, continuous, and time-variant. Indexing provides the means to search for similar motion in a mocap database, to recognize a query action as one among several motion classes, and fosters the re usability of motion data to generate animation automatically. The topic of my research is the design, implementation, and evaluation of several approaches to the motion indexing problem. We consider three dierent existing types of whole-body motion indexing algorithms: dimensionality reduction based techniques, feature function based techniques, and dynamic time warping based techniques. Theadvantages and disadvantages of these techniques are explored. We evaluate the performance of each technique using a subset of the CMU Motion Capture Database and its corresponding annotation. These experimental results will allow for an objective comparison between the different indexing techniques and for assessing the deficiencies of whole-body indexing techniques.