Comparison Of Human Action Recognitions In Monocular Videos Using Dtw And Hmm

In this paper, the use of two well-known recognition algorithms which are Dynamic Time Warping (DTW) and Hidden Markov Model (HMM) are studied and compared. From each frame in monocular videos, we first estimate the 3D human pose which consists of 3D coordinates of specific human joints using an efficient 3D human modeling technique; then convert them into a set of geometrical relational features (GRF), which describe the geometric relations among body joints of a pose for dimensionality reduction and discrimination increase. Next, the k-means clustering technique is applied to those GRFs to generate feature vectors for further dimensionality reduction. Finally, we use DTW and HMM in succession for recognition of actions and then compare their recognition performance. In our system, in order to recognize the repeated actions we use a variation of original HMM which is cyclic HMM (CHMM). The experiment results on IXMAS dataset show that CHMM stands out DTW in terms of recognition rate