Preprocessing and Normalization of 3D-Skeleton-Data for Human Motion Recognition

Abstract

One key task for motion recognition using machine learning algorithms is the preprocessing of the input data. In this work 3D-skeleton-data is used to classify 19 motion exercises. Due to different body shapes and deviations in the movements from different subjects it becomes necessary to normalize the data. This work addresses the normalization of 3D-skeleton-data by indicating body joint angles and normalization to an independent coordinate system. The recognition is based on a Support Vector Machine (SVM) and is evaluated on a dataset containing examples from 21 subjects. The recognition accuracies using different normalized feature combinations are examined. The authors conclude that joint angles are best suitable for the recognition of motion exercises.