Pseudo-Coloring on Dual Motion History Image for Action Recognition from Depth Maps

Depth video-assisted human action recognition has gained huge research interest due to its numerous advantages over traditional RGB Videos. Hence, this paper considers depth action video sequences as input data and proposes a new 2D-depth action descriptor called Dual Motion History Image (DMHI). DMHI is an extended variant of MHI which encodes each frame of action sequence with respect to its pre- and post-frames. Further, to enhance the visualization capability, a pseudo-color coding mechanism is introduced which can explore even minor variations in grayscale intensities with a wide range of color hues. Finally, for feature extraction and classification, we employed the standard pre-trained deep learning model, ResNet. Simulations on a small subset of the HuDaActRGBD dataset prove the effectiveness.