A novel human posture estimation using single depth image from Kinect v2 sensor

In this paper, we propose an approach to estimate general posture of the human-body. We perform various 2D scans of the body to carry out their Fast-Fourier-Transform(FFT) to extract features which can be fed to a 2-layer feed-forward neural-network. This approach can offer an effective procedure given limited resources which are usually available for deep learning approaches. In comparison with the literature, our proposed method doesn't require any specific skeleton points of the human body, results in a reduced level of computational complexities. We compared our method with the state-of-the-art and the results show an increased level of classification accuracy. The training dataset is captured from a single subject moving with various postures. If the test-data is also captured from the same subject, then the overall level of accuracy is 98.2% while if the test images are captured from different subjects then the accuracy of the estimation would be near 91.5%. We also test our method in the restricted field of view where parts of the subject's body are not in the field of sensor view, the result shows that in this case, the accuracy is as high as 79.1%.