Recognition Of Human Action More Somatosensory Equipment

With the successful promotion and widespread popularization of kinect, human action recognition based on multi-kinect has a wide scope of promising applications in many areas such as smart surveillance, advanced perceptual interfaces, motion analysis and so on. Human motion data fusion based on multi-kinect can solve the problem of incomplete data while tracking the moving target. In this paper, we analyzed the multi-kinect calibration, the description methods of human motion, motion data fusion, human action recognition methods and so on. Our contributions and innovations are summarized as follows:(1) Camera calibration method. The basic principles of monocular camera calibration based on pinhole camera model were introduced in detail. Then binocular camera imaging model was setted up. Finally a multi-kinect calibration algorithm based on the Zhang’s calibration method was proposed and implemented.(2) The methods of describing human motion. Firstly, the basic principles about how the motion data captured by kinect were briefly introduced. On the basis of it, a hierarchical skeleton model which used for describing human body topology was proposed. Then a calculation method which used for calculating the joint’s spatial orientation was proposed. At last a standard skeleton model in line with human body topology was modelled. And skeleton model’s constraints which based on the constraints of physical and motion were designed.(3) Human motion data fusion. Through analyzing kinect’s characteristic, a filtering method was proposed. It is human motion data filtering method based on kinect’s credibility tag. Then we proposed a data fusion algorithm based on joints’variable weight. Finally through analyzing human motion model, an improved kalman predictor was adopted to estimate the3D position of human joints. Experimental results prove that by adopting the algorithms proposed above, the system can present integrated skeleton model in each frame.(4) Human action recognition. Through analyzing the principle component analysis method, we implemented the reduction of the dimension of motion data and extracted the features. Then the basic principles of support vector machine were introduced and studied three kinds of SVM algorithms which were used for multi-class classification. Then a human action recognition method based on principle component analysis method and support vector machine method was proposed. Experimental results prove that by adopting the algorithms proposed above, the system can recognize different human action effectively.