Real-time Fall Detection Prototyping with YOLOv11 Using an Integrated Multi-dataset Framework

Abstract

Falls are a major public health issue among the elderly due to their serious health, psychological, and economic effects. This study proposes and implements a human fall detection system using the latest YOLOv11 architecture. Five public datasets were combined into an All-In-One (AIO) dataset to enhance visual diversity and model generalization. Data preprocessing involved normalization, re-annotation, and label harmonization into two classes: fall and no-fall. Experimental results show that YOLOv11n, with a learning rate of 0.001, achieved optimal performance (mAP@0.5 = 0.98, recall = 0.95, F1-score = 0.95) while maintaining an inference latency of 3.6 ms, enabling real-time detection. These results outperform YOLOv7 and YOLOv8, confirming the reliability of YOLOv11 in human fall detection. The trained model was deployed in a real-time desktop prototype integrated with a Telegram Bot for instant alerts, demonstrating practical potential in healthcare and eldercare applications. The study’s novelty lies in introducing a large and diverse AIO dataset, as well as a lightweight real-time prototype for fall detection.