View Invariant Spatio-Temporal Descriptor for Action Recognition From Skeleton Sequences

Recently, human action recognition based on 3-D skeleton sequences has gained significant interest in the computer vision field. However, skeleton sequences are sensitive to noises, viewpoint variations, and similar movements. To sort out these problems, this article proposes a view invariant spatio-temporal descriptor (VISTD) that encodes the positions of skeleton joints in each frame of an action sequence. VISTD is a combination of view invariant skeleton joint descriptor and spatio-temporal skeleton joint descriptor. Further, we employ a standard convolutional neural network (CNN) architecture for classification. To analyze the impact of fusion methods, the results of CNN model are fused with different fusion rules. Extensive simulation experiments were carried out over the proposed model through UWA3DII dataset and NTU-RGB+D dataset. The comparison of experimental results with existing methods explores the superiority.