An HMM based approach for video action recognition using motion trajectories

In this paper, we propose a new approach for video action recognition from motion trajectory data. Firstly, we extract motion trajectory features from trajectory groups. Then Hidden Markov Model is used for modelling different video actions. Secondly, we propose an improved parameter estimation algorithm for HMM. Compared to the other traditional HMM learning algorithms, our new method has several advantages. It avoids the problem of being tracked to local optimal. The proposed method is capable of leaving the local optimal and finding global optimal. In the learning stage, a set of HMMs are trained for different type of video actions. The trained HMMs are used for video action recognition in a later recognition stage. Experimental results on different sports actions show that our Evolve-HMM outperforms the traditional Baum-HMM algorithm.