Abstract:Aiming at the problem that the accuracy of behavior detection using Spatio-Temporal Graph Convolutional Network (ST-GCN) in the existing fall detection methods needs to be improved, and the time information is not enough utilized, a fall detection method based on lightweight YOLOv3 human target detection model combined with human skeletal feature points is proposed. In this method, the AlphaPose algorithm is used to obtain the information of human skeletal feature points in real time. On the basis, combined with the improved ST-GCN model, the enhanced behavioral spatio-temporal information is extracted, so as to detect falls more accurately. The test results on the general data set and the self-built data set show that the method is effective in fall detection.