Recognizing 3D actions with shallow convolutional neural networks and quad-joint relative volume maps

Abstract

In this study, we present a new set of feature maps for use with a Circular convolutional neural network (CCNN) that overcomes the shortcomings of the prior maps and allows for superior pattern discrimination. By leveraging the local relationships among joint movements represented by three-dimensional quadrilaterals constructed for every conceivable set of four joints, these novel characteristics calculate the volumes of these time-varying quadrilaterals. As a result, they produce color-coded images known as spatio temporal quad-joint relative volume maps (QjRVMs).