Research On Key Issues Of Video Based Weight Lifting Training System

Intelligent computation is one of objectives of science development. Thedevelopment of computer hardware and software technology and the development ofa few supporting disciplines such as image processing, computer vision, patternrecognition provide a solid foundation. If the intelligent technology could beintroduced into sport training, it is helpful to provide more scientific and effectivetraining.In this thesis, facing the video based weight lifting training system, I study theproblem of disc tracking and human motion tracking, and provide correspondingreasonable resolutions. The details are as follows:1. We proposed a point correspondence statistics based disc tracking approach.First we extract a few feature points using Harris feature point detector, and build therelationship between feature points and the centre of the disc. Then we utilize LucasKanada tracker to track these feature points. In the subsequent frames we choose thevalid feature points using a statistically optimal method and predict the center point ofthe disc using the valid points. Using the correct disc center the location of invalidpoints is corrected. This method can avoid the larger tracking error and provide robusttracking result.2. We proposed a foreground segmentation approach to combining codebookscheme and Belief Propagation (BP). We construct background model using codebookscheme. Then using time filtering we filter out the transiently appearing objects andintegrate the periodically moving objects into the background model. Finally, usingthe belief propagation optimal scheme based on Bayesian theory which could fill thehole in the segmented foreground.3. We implement the particle filtering based human motion tracking approach.Based on the framework of Bayesian particle filtering, we implement a scheme totrack human joints. Firstly I construct the initial appearance model of human byinitializing the human pose state vector. The color histogram of different humancomponents is obtained to describe the appearance model. Furthermore, we resolvethe problem of obstacle by ordered matching and symmetric constraints. Finally,considering the similarity of weightlifting, we introduce a motion model to guide thetracking algorithm which improves the accuracy of tracking result.