An Early 3D Human Action Recognition (3DHAR) System by Training Subactions of an Action using Deep Networks

Abstract

The Human action (HA) comprises of a set continues sub-actions. In this way, our point is to foresee the HA's utilizing those arrangement of sub-actions before finishing an action. Usually, human action recognition (HAR) has been performed utilizing joints, RGB recordings and Depth in all these cases forecasting an incomplete HA became most difficult task. So as to beat this issue, we proposed a novel approach utilizing bidirectional LSTM system so as to characterize the an action from a fully trained convolution neural network (CNN) features. CNN features are acquired from the 3D skeleton information modelled to joint angular displacement maps (JADM). JADM is spatio temporal presenation of a 3D skeleton information and is taken from our past works. This proposed strategy 3DHAR-Net is equipped for dealing with enormous action sequences to predict the action classes by analysing series of sub-actions in time sequences. The entire experimentation is carried on our self-made 3D human activity datasets KLYoga3D and KLHA3D-102 action dataset and accomplished better outcomes and our method is further evaluated on the three benchmark datasets CMU, HDM05 and G3D.