Action Recognition Based On The Probability Graph Model

With the rapid development of computer technology, intelligent equipment functions become more perfect, computer vision technology has been rapid development, produced a lot of research results, including intelligent video surveillance, is widely used in daily life, accurately recognition the action becomes the key to the application. Dynamic Bayesian network is one of the most typical probability graph model, it can catch dynamic relationship between random variables,and the human action is a process of dynamic, so, in this paper, dynamic Bayesian network is used for action recognition, mainly to do the work in following two aspects.The first work is human action recognition based on Dynamic Bayesian network,the database is MSR Action3 D human action database. The database extraction the features of position characteristics and movement characteristics, and for normalization, PCA dimension reduction, we get the data required for training process. Then these characteristic data as the observation data, the two kinds of features were seen as a Gaussian mixture model, to establish a dynamic Bayesian network, parameter procedure is followed, parameter learning algorithm method used is the EM algorithm, the method can be more accurate estimation parameters, thereby improving the recognition rate of the classification process. In the test of cross subjection, the recognition rate is higher than the other references, and the complex action data sets AS3 is the recognition rate also improved.The second work is initialized based on the training data used in action recognition, which does not require repetitive training to find a suitable model parameters, thus greatly reducing training time action to solve the time-consuming problem of random problem. Firstly, the observed data are separated by time, and then calculate the average value of each sub-sequence, this average is initialized centers of K-means algorithm, the parameters of Gaussian mixture components is then initialized, and then parameter learning, probabilistic inference, to complete thewhole process of action recognition. Compared to the random initialization method improves the recognition rate, and for complex action set AS3, the highest recognition rate of random initialization can find, can be achieved the balanced of the recognition rate and time, better meet the needs of practical applications.