Fall Detection in Elderly People at Home Using the YOLOv11-Pose Model

Abstract

Falls are a leading cause of serious harm and unintentional death among the elderly, especially in homes where direct monitoring or rapid assistance is unavailable. Most of the models based on computer vision developed to determine falls lack decent performance in the real world, e.g., they cannot work properly in the case of views lighting conditions. Such deficiencies result in the realization of detections or false alarms particularly in residences. The study will focus on promoting the accuracy of fall detection among older adults, by proposing a strong real-time system capability, where object detection is combined with human pose estimation. In this study, we propose an automated fall detection system using computer vision techniques, specifically the YOLOv11-pose model. A phone camera was used to capture high-resolution image data, which was then properly processed, labelled, and optimized before training multiple versions of YOLO11-pose using standardized hyperparameter settings to ensure fair comparison. Experimental results showed that all models performed well in fall detection and pose estimation, with smaller models ideal for rapid implementation, achieving high frames per second (FPS) of around 95 fps and larger models, such as YOLOv11m, achieved an accurate mAP@50 of 99.43%. This study determined that YOLOv11-Pose is an effective model for detecting falls in the elderly, with promising accuracy and speed. Future directions include improving performance in challenging imaging conditions, expanding the model to include more activities.