Real-Time Vision-Based Risk of Fall Detection Using YOLOv8
Sherif Abdelfattah, Faraz Ali, Sai Gowtham Kalleti, Mohamed Baza, Mahmoud M. Badr · 2025
Falls among patients lying in bed continue to be a serious safety concern in hospitals and care facilities, often leading to injuries and increased healthcare costs. Traditional solutions, such as wearable sensors or pressure-based systems, frequently suffer from low user compliance and generate unnecessary alarms. To address these limitations, this paper presents a real-time vision-based fall detection system that uses a unified YOLOv8 framework for both bed segmentation and human pose estimation. The system identifies potential fall scenarios by analyzing how a person’s posture spatially relates to the edges of the bed. Using a combination of real-world images and synthetically generated images, the system extracts body keypoints and bed corners to compute spatial features. These features are then used to classify each situation as a fall or non-fall using machine learning models. Among the models tested, XGBoost demonstrated the highest accuracy, achieving strong performance in detecting falls. The system is also efficient, capable of running at high frame rates on edge devices, making it suitable for real-time deployment in clinical environments. The results confirm that the proposed method is not only accurate and fast, but also practical for everyday use. It offers a reliable and scalable solution that can improve patient safety without requiring physical contact or user intervention.