| As an important way of information transmission, the theoretical research value of human action recognition is very important. It has a broad application prospect in the smart home area, intelligent monitoring, video games and so on. Especially in the smart home area, everyone wants to be more intuitive when operating household appliances. The premise of the interaction realized is that the computer can understand the intent of the person through observing the action. Initially, the study of action recognition is based on two-dimensional image. But because of lacking one-dimensional distance information, the feature extraction and representation have been limited. It is difficult for us to dig out more details. As a result, some researchers create 3D reconstruction based on 2D image, but the modeling process is very complex. With the appearance of depth sensor, the 3D scene information can be obtained easily by combining the depth image and common optical image, and avoid the complexity of 3D reconstruction. Based on the above analysis, this paper choose Kinect depth sensor which providing the function of skeleton positioning. In this paper, we research and implement the human action recognition algorithm which is suitable for home.Based on the study of SDK function and the skeleton tracking function of Kinect, the 3D data of the joints was obtained by using API. Based on the data, we construct human skeleton model as the basis of the following research.Creating feature model based on the human skeleton is the key to the identification of human action. In this paper, we build feature vectors based on the characteristic of human body structure. Then using the combination of vector angle and modulus ratio to describe the human motion. After that, we validate the model by using the data set. The study of the action recognition algorithm is the focus of this paper. In this paper, the SVM algorithm is used for motion recognition. This paper studied the principle of SVM. Then the methods of parameters selection and optimization are given. Last, simulated on three optimization methods. Though camparison of the three methods, GA method have the fast search speed.This paper also designed a feature collection system and designed a set of action based on the application environment. Then applied SVM algorithm to action recognition based on the feature extraction, normalization, principal component analysis, parameter determination and training process. Last, the paper analyzed the classification results, reaching the accurate classification of action. |
