Real-Time Fall Detection for the Elderly in Home Settings Using a Vision-Based YOLOv8-Pose Model

Source abstract

Fall constitute a significant source of death and morbidity among the elderly, particularly in domestic settings when quick assistance is unavailable. Accurate and timely detection of falls is crucial to prevent associated health repercussions. The study proposes a deep learning-based fall detection model YOLOv8-pose that features simultaneous fall detection and pose estimation. A new dataset was created, including 186 high-resolution videos (1080p, 30 fps) captured by a smartphone camera in the home-like environment simulating falling and everyday activity. The video frames were marked and organized according to the COCO format to attain effective training. We used transfer learning to train the state-of-the-art YOLOv8-Pose model in order to detect falls. The system also analyses human skeletal keypoints in real time to accurately differentiate between fall events and non-fall activities, resulting in strong detection performance and a low false positive rate. An automatic alarm system was implemented to deliver immediate fall alerts to providers using the Telegram messaging bot. The proposed approach is an economical, non-invasive, and highly effective alternative for continuous monitoring of the elderly in a household environment, potentially enhancing safety and reaction time during emergencies. The experimental findings revealed that all of the models did a good job of detecting falls and estimating poses. Larger models, such YOLOv8m-pose, got a 99% accurate mAP@50 and the confusion matrix showed that the model was able to classify falls well, with a very low error rate in classifying other cases, reflecting its high effectiveness in distinguishing between falls, normal falls, and bending. This study found that YOLOv8-Pose is a good model for finding falls in older people since it is fast and accurate. In the future, we want to make the model better at imaging under difficult situations and broadening the model to incorporate more activities.