Vision-based fall detection system for the elderly using image processing and deep learning
Đức Nguyễn Tiến, Việt Đỗ Hoàng, Thai Nguyen Van · 2023
Population aging is happening in developed and some developing countries, this means the proportion of elderly people in society is increasing dramatically. The elderly often suffers from bone and joint diseases that make their daily live difficult. Therefore, accidental falls are a major cause of loss of autonomy, injuries among the elderly. Healthcare surveillance systems need to be improved to take care of these elderly due to the lack of nurses, this paper represents a solution to this problem. By using a vision-based architecture, the healthcare surveillance system can detect fall accidents in people's daily life activities and then notify the nurse to have in-time assistance. Yolov3-tiny is applied to detect humans in the frame, bounding boxes are generated to visualize the detection, then humans are tracked using the Kalman filter algorithm. AlphaPose is applied to generate key points from detected persons. Each keypoint coordinate will change frame after frame when persons in the frame are moving, then keypoint coordinate states are fed into ST-GCN as input, and the ST-GCN model will predict the probability of the person's activity (fall or not fall). Our model is more effective at detecting and predicting falls, distinguishing it from many other proposed models in the past. Some of our improvement include fall detection in low lighting conditions, multi-person fall detection, and fall detection when the human body is partially occluded. The experimental results show that our proposed vision-based surveillance system achieves high accuracy of 99.08%, precision of 98.84%, and recall of 98.03%.