Recurrent Neural Network based Action Recognition from 3D Skeleton Data

In this paper, we present an approach for human action recognition from 3D skeleton data. The proposed method utilizes Recurrent Neural Network (RNN) and Long Short Term Memory (LSTM) to learn the temporal dependency between joints' positions. The proposed architecture uses a hierarchical scheme for aggregating the learned responses of various RNN units. We demonstrate the effectiveness of using only a few joints as opposed to all the available joints' position for action recognition. The proposed approach is evaluated on well-known publicly available MSR-Action3D dataset.