Action Recognition System Based On Kinect

As an important category of human computer interaction,human action recognition has attracted more and more attention in recent years.Human action recognition has been applied in many aspects such as games,medical,sign language recognition,education and so on.In particular,as one of the most representative depth camera,Kinect brings unlimited vitality and challenges to this field.Different from the traditional RGB camera,Kinect can capture the RGB data,depth data and skeleton data,which make it possible to precisely recognize human action.This paper briefly introduces the common algorithms for human action recognition,Kinect and the database which used in this paper.On the basis of the traditional features,we combine two different skeleton features as a new feature and put it into the classifier.Different from the traditional Hidden Markov Model(HMM)in action recognition.We propose a posture selection method based on two-layer Affinity Propagation(AP),and apply it into HMM.In the aspect of motion recognition software platform,we design and implement a MFC based action recognition software system based on the mixed-language compiler of Visual Studio and MATLAB.This software can recognize the skeletal data acquired from the Kinect in real time.The main work of this paper can be summarized as follows:Firstly,we study the basic algorithms of human action recognition and the characteristics of skeletal data.On the basis of traditional skeletal features,we group two different skeleton features named the pair-wise relative position features and the skeletal angle features respectively.After a large number of experiments,we select the appropriate dimensions of the pair-wise relative position features and the skeletal angle features,and put these features into the classifier.Secondly,we improve the initialization process of the HMM.We propose a posture selection method named two-layer AP.The method selects the postures as the hidden states in HMM.We perform simulation on two databases MSR Action3 D and UTKinect.The results show that the method not only avoids the unstable factors in both K-Means and HMM parameter initialization,but also the postures selected by the two-layer AP can express the sequences of the action well.We can use these postures to represent the hidden states of HMM,so that we can get a higher recognition rateThirdly,we apply the proposed algorithm to the development of a real action recognition system.We design a MFC interface system which can obtain the skeleton data from the Kinect and recognize the human action in real time.The system not only can be used as a Kinect skeleton data acquisition system,but also can be used to accurately identify all kinds of training actions.