Research Of Human Motion Tracking Of Virtual Exercise Machine

The human-computer interaction based on computer vision has become an importantissue for research in the field of image understanding and computer vision. The virtualexercise machine based on computer vision is a portable, scalable mode of intelligenthuman-computer interaction.The planning of virtual exercise machine based on computer vision should include thedefinition of action, the tracking of human motion,3D reconstruction, and the output of actionmessage. Considering of the workload, this article, therefore, only make arelatively deepresearch on the tracking of human motion, the key part of virtual exercise machine. Theachievements are as follows:Firstly, the extraction algorithm of moving human is proposed. As for backgroundsubtraction based on the Single Gaussian Background Model, the idea of C-means clusteringis used to initialize the background model, to enhance the quality of the background and toreduce the impact of the first frame to the background. As for the binary foreground image, inorder to get a relatively complete moving human, the paper use the colorimetric theory toeliminate the shadow, and morphological filter are used to reduce noise and connect sectors.Secondly, the extraction algorithm of human joint is proposed.Traditional method ofextraction algorithm of human joint is to extract joint with marker by manual work, but in thispaper, the automatic extraction algorithm without manual work has been proposed. In thisalgorithm the trigger positions used to find jointsare obtained through the judgment ofextreme points, and the initial positions of joints are located through the Mean-Shift algorithm.Then by combination of the initial position, assumptions of marker’s information andthe Mean-Shift algorithm, the positions of joints are roughly located and through backgroundinhibit histogram and the Mean-Shift algorithm, the positions of joints are accuratelylocated.At last, the bandwidth of joints is determined by using Bhatacharyya distance anddichotomy. After the judgment of extreme points, the determination of initial position, roughlocation,accurate location, and bandwidth determination, the centers andbandwidth of thehuman joints are gained and tracking target templates are eventually established.Thirdly, the target tracking algorithm based on the Mean-Shift is improved. In order to overcome the difficulty of the traditional Mean-Shift algorithm in tracking fast targetand swerved target, the position of tracking target in the next frame is predicted throughimproved Kalman filter. To settle the problem of change of target size, the method of adaptiveupdate of bandwidth of kernel function based on Bhatacharyya distance is used. To solve theproblem that the target template need reflect the latest features of the target, the method oftarget template update based on Bhatacharyya distance is used.Finally, not only a planning for the virtual exercise machine is made, but the function oftracing the moving human automatically is realized.