Body Motion Analysis for Similarity Retrieval of Motion Data and Its Evaluation

Abstract

With the spread of motion capture systems, body motion data has been obtained without effort and quantity of its data has been increasing. We need an efficient method for target motion retrieval. In this paper, we propose the method of analyzing motion data in advance, and express the construction of movement with indexed labels for reducing similar-motion retrieval. The posture of each frame is distinguished into the pattern of the indexed labels, and we therefore try to index each motion data. Firstly, we reduce the number of dimensions of motion data with PCA (Principal Components Analysis) and do clustering. Next, a trained dictionary is consisted of the indexed motion data. Finally, SVM (Support Vector Machine) distinguishes the test data by using the dictionary. In this research, we made the trained dictionary from the movement of radio gymnastics, and then carried out an evaluation of discrimination by using break dance movement as test data. As a result, we could use the motion data with indexes, and obtained a good retrieval result in this experiment.