In order to accurately identify the fall posture of the elderly, an improved YOLOv8s fall detection model is proposed. Firstly, the SE attention mechanism module is introduced into the backbone network of the YOLOv8s model, which divides the channel features into multiple subgraph features, and fuses the features of different groups, so that the network can adaptively focus on the key features, suppresses the features that contribute less to the current task, and improves the feature extraction ability; secondly, the CIoU loss function is replaced by EIoU to accelerate the convergence rate and improve the accuracy and stability of the model. Finally, the trained model is verified on URFD+ and other data sets. The experimental results show that the precision rate of the model reaches 99.50%, the recall rate reaches 99.00%, and the mAP50 reaches 99.50%, which is better than the original model. Compared with the YOLOv5s+K-means++ model, the accuracy is increased by 3.22%, the recall rate is increased by 5.32%, and the mAP50 is increased by 2.38%. Compared with the C2D-YOLO model, the accuracy is increased by 10.00%, the recall rate is increased by 11.40%, and the mAP50 is increased by 7.80%. Compared with the YOLOv5s+C3new model, the accuracy is increased by 2.50%, the recall rate is increased by 6.80%, and the mAP50 is increased by 4.1%. The improved model has greater advantages than the original model and the current advanced model.