Video-Based Fall Detection Based on A Spatio-temporal Graph Convolutional Auto-Encoder

To achieve video-based fall detection, a spatiotemporal graph convolutional auto-encoder was designed based on human skeleton information and principles of currently available auto-encoders to extract the features of body posture. Then, considering falling as an abnormal movement, a video-based falldetection method was designed using the detection framework of abnormal movement. Thus, unsupervised automatic fall detection was achieved in this study. The test results on an open testing dataset indicated that the proposed method performs well in terms of both detection accuracy and time.