<p>Currently, mainstream fall detection technology is computer vision. This method is divided into two steps: detecting keypoints and detecting falling behavior. The commonly used keypoint detection algorithm is OpenPose, whose backbone is highly redundant and large, resulting in a low frame rate and poor deployment ability of the terminal. To address this issue, we propose the keypoint detection algorithm MossPose (Mobile SE and SA Attention OpenPose), which is composed of OpenPose, MobileNet, and attention modules. YOLOv5 is added for object detection before keypoint detection, which solves the problems of poor deployment performance and low accuracy after lightweight on mobile devices. The commonly used fall detection algorithms do not always consider the running speed and accuracy at the same time, therefore, we used the efficient and lightweight MobileNetv2 to detect falling behavior and calculate the aspect ratio of the external rectangle of the human body and the falling speed of some keypoints to determine whether falling behavior occurs. Experiments were conducted on the COCO2017 keypoints dataset, Le2i fall dataset, and UR fall dataset. The AP on the COCO2017 keypoints dataset was 53.3%, and by improving the human keypoints detection, the fall detection algorithm was optimized, the accuracies of the proposed method are 90.05% and 91.43% with 22 FPS on the Le2i and UR datasets, respectively, indicating the high accuracy of real-time detection.</p>