Novel human action recognition in RGB-D videos based on powerful view invariant features technique

Abstract

Human action recognition is one of the important topic in nowadays research. It is obstructed by several factors, among them we can enumerate: the variation of shapes and postures of a human been, the time and memory space need to capture, store, label and process those images. In addition, recognize a human action from different view point is challenging due to the big amount of variation in each view, one possible solution of mentioned problem is to study different preferential View-invariant features sturdy enough to view variation. Our focus on this paper will be to solve mentioned problem by learning view shared and view specific features applying innovative deep models known as a novel sample-affinity matrix (SAM), able to give a good measurement of the similarities among video samples in different camera views. This will also lead to precisely adjust transmission between views and study more informative shared features involve in cross-view actions classification. In addition, we are proposing in this paper a novel view invariant features algorithm, which will give us a better understanding of the internal processing of our project. We have demonstrated through a series of experiment apply on NUMA and IXMAS (multiple camera view video dataset) that our method out performs state-of-the-art-methods.