Recognition of human motion exercises using skeleton data and SVM for rehabilitative purposes

This work addresses the recognition of 19 gymnastic motion exercises for a training assistant. The system captures skeleton data in form of 3 dimensional body positions with the Microsoft™ Kinect v2. In the preprocessing phase, the 3 dimensional positions are transformed to joint angles. The recognition is based on a Support Vector Machine (SVM) with a polynomial kernel. Six subjects are used to train the models with 19 motion exercises. The models are tested with 15 different subjects afterwards. With the proposed method an accuracy of 81% correct classified motion sequences is achieved. From this work it can be concluded that autonomous systems for the recognition of motion exercises are a promising tool for rehabilitative purposes.