Motion evaluation by means of joint filtering for assisted physical therapy

The supervision of rehabilitation exercises is crucial for a successful therapy. Due to a lack of therapists, technical assistance systems have recently come into focus to assist patients during their exercises. Latest research proved that characteristic motion errors can be detected by using the Kinect skeleton joints in connection with Incremental Dynamic Time Warping (IDTW) and machine learning. However, the processed joints were manually selected and the classifier predicts in a frame-wise manner. In order to facilitate an extension with more exercises, a central issue of this paper is to realize an automatic joint selection with optimal classification accuracy. Moreover, we propose an algorithm that post processes the frame-wise prediction. The results for both joint selection and post processing are of high quality and therefore make a significant contribution to an efficient, perceptible and user-friendly feedback generation.