To solve the problem that existing graph convolution network (GCN) need to pre-define human skeleton topology and the model is large, a fall detection algorithm based on spatiotemporal adaptive graph convolutional network (ST-AGCN) is proposed. The network consists of three parts: firstly, HRNet, a human pose estimation algorithm, is used to extract human skeleton points from video and preprocess them into four-dimensional tensor. Secondly, the normalized embedded Gaussian function is introduced to obtain the human body topology by learning (without manual pre-definition), and the human body correlation features are obtained by spatial adaptive graph convolution. Thirdly, multi-scale convolution is used to extract temporal motion features to improve the model′s ability to obtain dynamic information. Simulations are carried out on public and self-built dataset, and the accuracy rates are 95.45% and 99.55%, respectively. The results show that the proposed algorithm is better than the current GCN methods, and the number of parameters is only a quarter of the latter, or even less. Another advantage of our algorithm is that it can be applied to different datasets.