The Recognition Of Action And Spontaneous Expression Based On Deep Neural Network

Aging of population and a large number of the disabled have become a serious social problem in the global scope, and developing service robot emphatically is an import method to solve this problem. Therefore, the investment in service robot is increasing and a variety of home service robots have been developed. However, there is no service robot can provide high-quality human-friendly service, the main reason is that robot doesn’t have the ability of independent learning. The premise of independent learning is the understanding of human behavior and intention. Robot’s understanding of human behavior and intention include two aspects:one is the understanding of human body movement; the other is the understanding of facial expression. The two aspects are the important research contents of national fund project-service robot’s independent learning of personification behavior under the support of smart space. This paper aims at the study of human action recognition and spontaneous facial expression recognition.As deep belief network can not deal with the temporal data, this paper proposes temporal deep belief network according to conditional restricted Boltzmann machines. The action data have time correlation, then the frames before and after the current frame can provide contextual information. Therefore, the modified model adds the time information and improves the accuracy of action recognition, especially for the actions with similar poses. In the same time, the proposed model doesn’t need segment action sequences manually and can recognize one action from any time. It is the true realization of real-time recognition, and makes a good foundation for the practical application.Compared to posed expression, the recognition of spontaneous expression is more effective in practical application. But the recognition of spontaneous expression is more difficult for its own characteristics, such as unobvious features, unpredictable emergence and unobtainable database. The current methods are largely depending on the mark of feature points. To solve the above problems, for the first time this paper applies convolutional neural network to the recognition of spontaneous expression taking the advantage of the powerful learning and expressing ability of convolutional neural network. Experiments show that convolutional neural network can also get comparable results without feature points. In addition, this paper testifies that pre-training is effective using posed expression aiming the problem of unobtainable spontaneous expression database.