Human motion classification and management based on mocap data analysis

Abstract

Human motion understanding based on motion capture (mocap) data is investigated. Recent rapid developments and applications of mocap systems have resulted in a large corpus of mocap sequences, and an automated annotation technique that can classify basic motion types into multiple categories is needed. A novel technique for automated mocap data classification is developed in this work. Specifically, we adopt the tree-structured vector quantization (TSVQ) method to approximate human poses by codewords and approximate the dynamics of mocap sequences by a codeword sequence. To classify mocap data into different categories, we consider three approaches: 1) the spatial domain approach based on the histogram of codewords, 2) the spatial-time domain approach via codeword sequence matching, and 3) a decision fusion approach. We test the proposed algorithm on the CMU mocap database using the n-fold cross validation procedure and obtain a correct classification rate of 97%.