Skeleton geometric transformation for human action recognition using convolutional neural networks

In this paper we present a methodology for understanding human actions. We try to compensate for viewpoint changes, by applying geometric transformations to 3D skeletal joint information. More specifically, motion information regarding human skeletal joints is pre-processed to create 2D image representations. Then a DST transformation is applied, to transform them to the spectral domain. Convolutional Neural Networks are then used for classification. We evaluate our approach in actions that may be used in an ambient assisted living scenario.