A quick filtering for similarity queries in motion capture databases

Abstract

A similarity retrieval of motion capture data has received substantial attention in recent years. In this paper, we focus on feature extraction and quick filtering methods in the similarity retrieval system. A representation of motion capture data is joint angles, which can distinguish different human body poses. We propose a new technique for dimensionality reduction based the average and variance of joint angles. Our dimensionality reduction is simple to understand and implement. In experiments, twenty dance motion clips each of which is different in length and style, are used in the test data set with a total of 60,000 frames. The results of our quick filtering show an achievement on the recall and precision up to 100% and 70%, respectively. © 2009 Springer-Verlag Berlin Heidelberg.