3D Human Activity Recognition

Abstract

Human action recognition based on 3D skeleton data has received considerable research attention. Existing methods for modelling skeletons which rely on hand-crafted features or deep learning fail to capture the long-term temporal information and the complex spatial structures which are very important to predict human action. In this work, we applied an existing model called Spatial-Temporal Graph Convolutional Networks (ST-GCN) using a different dataset (NTU-RGB 120).In addition, we apply recent findings in graph neural network (bottleneck problem) by changing the network architecture of ST-GCN(reducing the number of the network layers), and we obtain better recognition performance. We also add a code to compute the confusion matrix to evaluate the performance of ST-GCN. This model tackles the challenges faced by previous methods by automatically learning the spatial and temporal information of the skeleton. Furthermore, this model is the first work that applies Graph Convolutional Networks (GCNs) to model skeleton data as spatiotemporal graphs where nodes correspond to human body joins and edges correspond to connectivity between joints. Experimental results on NTU-RGB 120 dataset shows the effectiveness of the proposed model. We compared the performance of this model with a Two-Stream ST-GCN model, and with a new model with less number of layers.