Aiming at the problem that the current human fall detection methods can not meet the high accuracy and detection speed at the same time, an improved YOLOv5 algorithm is proposed to detect fall or not. This method adds ECA to the backbone of YOLOv5s to improve the ability of the model to extract features, and increases precision without too much computation. PANet is modified to BiFPN for the weighted fusion of features of different dimensions extracted by network of neck. Experiments on the Le2i fall detection public fall dataset shows that the precision of the improved method has increased by 2.7 percentage points, and has a certain improvement compared with other mainstream algorithms, which proves the effectiveness of the algorithm.