| E. Mathe, A. Maniatis, E. Spyrou, Ph. Mylonas |
| A Deep Learning Approach for Human Action Recognition Using Skeletal Information |
| Advances in Experimental Medicine and Biology, Volume 1194, Springer Nature, Switzerland |
|
ABSTRACT
|
| In this paper, we propose a novel visual representation of human actions, based on the discrete Fourier transformation (DFT). More specifically, we concatenate raw signal images that result from the 3D motion of human skeletal joints. The input required for the extraction of these joints consists of aligned RGB and depth video sequences. It is performed using the well-known Kinect v2 camera and its accompanying SDK. Moreover, we propose a novel convolutional neural network (CNN) architecture which uses as input the DFT transformation of the aforementioned images. We evaluate the proposed approach using the challenging PKU-MMD dataset consisting of 51 human actions, and we demonstrate that the proposed approach may be used in real-like environments for the recognition of activities of daily living (ADLs).
|
| 26 June , 2020 |
| E. Mathe, A. Maniatis, E. Spyrou, Ph. Mylonas, "A Deep Learning Approach for Human Action Recognition Using Skeletal Information", Advances in Experimental Medicine and Biology, Volume 1194, Springer Nature, Switzerland |
[  PDF] [
BibTex] [
Print] [
 Back] |
