Proceedings of the 31st International BCS Human Computer Interaction Conference (HCI 2017) (HCI)
digital make-believe, with delegates considering our expansive
3 - 6 July 2017
This study uses machine learning methods to analyse Kinect body gestures involved in the user interaction with exergaming systems designed for physical rehabilitation. We propose a method to improve gesture recognition accuracy and motion analysis, by extracting from the full body motion data recorded by the Kinect sensor three important features which are relevant to physical therapy exercises: body posture, movement trajectory and range of motion. By applying the Hidden Markov Model (HMM) and Dynamic Time Warping (DTW) algorithms, we obtained an improved accuracy by selecting specific features from the public UTD-MHAD full body gestures database (with up to 56% for HMM and 32% for DTW). Preliminary results show a positive correlation between the movement amplitude and the envelope feature (r = 0.92). Thus, this approach has the potential to improve gesture recognition accuracy and provide user feedback on how to improve the movement performed, in particular the movement amplitude. We propose further improvements and method validations to be the basis of creating an intelligent virtual rehabilitation assistant.