Artificial Increase of 3D-Skeleton-Data for Human Motion Recognition using supervised SVM and NN
Jan Paul Vox, Frank Wallhoff · 2018
This work addresses the recognition of gymnastic human motion exercises. For the classification, Neural Networks (NN) and Support Vector Machines (SVM) are being used. They show similar accuracy results for 20 different motion exercises. The algorithms are trained in a supervised manner with data from only one person. In order to achieve a unified classification framework for different subjects with different body sizes and personal deviations in the motion sequences, the training data was artificially increased by introducing deviations. More concrete, additional data was generated and added to the training set by random interpolation in a parameterized way. The recognition was evaluated with 20 unknown subjects. The authors concluded that an accuracy of 87% can be achieved.