A comparison of features extraction method for HMM-based motion recognition

The HMM-based human motion recognition on has recently gained lot of attention. In this paper, we research motion recognition based on joint angle trajectories derived from VICON System. The purpose of this paper is to find a better features extraction method in motion recognition system, even if only limited amount of training data is available. We achieve this purpose by significantly reducing the amount of input features. We have seen that human motions display only a few independent degrees of freedom (DOF) during resent research. We compared the feature extraction method, Brute-Force Feature Selection (BFS), Sequential Forward Selection (SFS) and Linear Discriminate Analysis (LDA). The experimental results show that when we reduce the number of features up to 3, we could get better human motion recognition performance.