The Research On Action Recognition Algorithms Based On Hidden Markov Template Model In Video Sequences

Recently, action recognition and behavior understanding is an active topic in computer vision field, which is adapted to many fields such as human-machine interaction, Intelligent video surveillance, virtual reality, motion analysis and other fields. The research of this topic has great value in academic study and practical application. Because of non-rigid nature of human movement, the variations of inter-class are large while the extra-class variations are small. So it’s more difficult to detect and recognize human action than other objects from video.At present, action recognition algorithms based on Hidden Markov model have the following problems:(a) It’s not easy to construct the probability model between observations and hidden state correctly;(b) Each observation corresponds to a group of local features, establish the relationship between the local features is very important,(c) In the machine learning stage, sometimes we need to label the important frames manually to get better results,(d) The process of inference is time consuming. According to these problems above, this thesis present a Hidden Markov Template (HMT) model for recognizing human activities from video clips. The main content can be summarized as follows:The basic structure of this model is consists of a series of hidden states and observations. This novel uses active template model to compute the probabilities between observations and hidden state. Compared with the rigid templates, active templates are robust to the changes of perspective and the objects’appearance. So the active template model can accurately construct the probability model between observations and hidden state. We can get the transition probability of key pose by statistical method. Transition probability of Hidden Markov Model can describe the relationship of key Poses correctly. Given a set of training video of an activity class, we can learn the model parameters automatically with a modified version of the Expectation-Maximization (EM) algorithm. The inference can be solved efficiently by Dynamic Programming. In the platform of MATLAB, this novel tests the present algorithm on Weizmann dataset. Experimental results demonstrate that "left to right" hidden Markov model has obvious advantages in describing the switching rules of key poses. Furthermore, extracting the combined features can achieve the best recognition results. Each action can be divided into several poses, increasing the number of pose can improve the recognition accuracy, however, when the number of pose reaches four, and then increasing the number has little effect on the final result of attitude. Compared with other state-of-the-art methods, the algorithm we present can get higher recognition accuracy and doesn’t depend on the result of background segmentation.