Human Fall Detection Using Spatial Temporal Graph Convolutional Networks.

Abstract

Falls are a serious issue in society and have become a major topic in the healthcare domain. Because of the rapidly increasing number of elderly people, falling can cause serious consequences for the elderly, especially if the fallen person is unable to get up. Early detection of falls and reducing waiting times help save the lives of the elderly. The increasing number of cameras in our daily environment, coupled with the presence of a smart environment, makes the vision-based system the optimal solution for fall detection tasks. A vision-based system using convolutional neural networks (CNN) to detect a fall event in different scenes with different background models is proposed in this paper. For privacy concerns and to avoid complex background problems, skeleton data for the human body was used as an input to the network. A pre-trained spatial temporal graph convolutional network (ST-GCN) model is used for the fall event classification task. ST-GCN classifies the extracted spatial and temporal features from the skeleton data of a detected human as falling or non-falling. To evaluate the proposed system, three public datasets (FDD, URFD, and MCF) that have different environmental issues are used. The experimental results prove the efficiency and robustness of the proposed system in complex situations. The proposed system achieves high performance rates compared to several state-of-the-art systems, with an overall accuracy of 98.6%.