Human Action Recognition (HAR) gained huge research prospect due to its widespread applicability in different applications related to Human computer interaction, Human Robot Interactions and Visual surveillance etc. However, the traditional RGB videos assisted HAR has poor performance as they composed of different illuminations, viewpoints, clothing etc. The emergence of 3D skeleton sequences sort out these problems and shown a new direction for HAR. However, the skeleton sequences are sensitive to noises, and similar movements. Hence, this paper proposes a new Action Descriptor called as Spatio-Temporal Joint Descriptor (STJD) for action recognition from skeleton sequences. STJD encodes the both spatial and temporal movements of an action and ensures improved recognition accuracy especially for action with similar movements. Initially, STJD segments each frame of skeleton sequence into different local segments and each segment is encoded with Spatial Skeleton Joint Descriptor (SSJD). Further, the SSJDs of each frame are encoded with Temporal Skeleton Joint Descriptor (TSJD) and fed to a pre-trained deep learning model ImageNet for classification.

Human Action Recognition, Skeleton, Spatio-temporal information, Deep learning, F-Score.