Available at: https://digitalcommons.calpoly.edu/theses/1246
Date of Award
MS in Computer Science
The automatic assessment of human physical activity performance is useful for a number of beneficial systems including in-home rehabilitation monitoring systems and Reactive Virtual Trainers (RVTs). RVTs have the potential to replace expensive personal trainers to promote healthy activity and help teach correct form to prevent injury. Additionally, unobtrusive sensor technologies for human tracking, especially those that incorporate depth sensing such as Microsoft Kinect, have become effective, affordable, and commonplace.
The work of this thesis contributes towards the development of RVT systems by using RGB-D and tracked skeletal data collected with Microsoft Kinect to assess human performance of physical exercises. I collected data from eight volunteers performing three exercises: jumping jacks, arm circles, and arm curls. I labeled each exercise repetition as either correct or one or more of a select number of predefined erroneous forms. I trained a statistical model using the labeled samples and developed a system that recognizes specific structural and temporal errors in a test set of unlabeled samples. I obtained classification accuracies for multiple implementations and assess the effectiveness of the use of various features of the skeletal data as well as various prediction models.