Basic Action Recognition Based On Deep Learning And Hidden Markov Model

There is an improvement in living standards significantly because of the development of technology, and a great deal of manual operation can be done automatically by machines in our life. It is hoped that machines can communicate with mankind and provide valuable message directly. Consequently, computer vision and image processing have become the concernful research contemporarily. Action recognition, as one of the main research topics, find a wide utilization in human-computer interaction, virtual reality, intelligent monitoring, and video retrieval.Action recognition can be accomplished in two stages:training and testing by traditional technology. In the meantime of training stage, the characteristics of human action video clips is extracted firstly, then a classifier is trained by applying the characteristics, the two steps are entirely separated. In the meantime of testing stage, the category of unknown behavior is determined by characteristics and trained classifier. The exacted characteristics heavily influence the validity of classifying. As a deep learning model, convolution neural network can extracted abstract features from the gray image directly and the parameter of model is updating by classifying results. The paper came up with a cascaded model based on convolution neural network and hidden markov model for basic action recognition, which combines the advantage of traditional recognition methods and deep learning. The major works of the thesis can be concluded as follows.(1)The performance of the behaviors in situ and the behaviors not in situ shows a larger difference, hierarchical decision classifier was used for higher recognition accuracy. A movement scope threshold classifier was put forward to pre-sort behaviors into two categories:the behaviors in situ and the behaviors not in situ, which was designed by movement scope of foreground target binary image sequence.(2)A 3DCNN-HMM cascaded model was proposed for basic action recognition. Firstly, extracted features including gray, gradient and optical flow from video cube; and then trained a 3D convolution neural network model for extraction of abstract characteristic; finally, trained a hidden markov model by the abstract characteristic for basic action classifying.(3)In order to improve the efficiency of 3DCNN-HMM cascaded model, a 2DCNN-HMM cascaded model was put forward.2DCNN-HMM cascaded model adopts two-dimensional convolution kernels to extract features of each frame from video cube respectively, which keep all the dynamic information of actions, so optical flow information is needless. Then a two-dimensional convolution neural network was trained for features extraction and a hidden markov model was trained for basic action recognition.