A Deep Learning Approach To Human Fall Detection With YOLOv12 Versi Nano

Florensce Sumarauw (Sam Ratulangi University)
Febriyanti Ludja (Master Program of Informatics, Postgraduate Program, Sam Ratulangi University, Manado, Indonesia)
Robby Moody Lintong (Master Program of Informatics, Postgraduate Program, Sam Ratulangi University, Manado, Indonesia)

The International Journal of Informatics, Data, and Emerging Applications (IDEA International Journal)
Vol. 1 No. 2 (2026): International Journal of Informatics, Data, and Emerging Application

The growing elderly population faces a heightened risk of falls. Current monitoring solutions, such as in-person observation by caregivers or wearable sensors, suffer from several drawbacks. These include low accuracy, user discomfort, and a constant need for human oversight. To overcome these issues, automated approaches using Convolutional Neural Networks (CNNs), especially the YOLO algorithm, have been applied for fast and real-time fall detection. In this study, the original YOLOv12-nano model, which is known for its lightweight architecture, low latency, and efficiency on edge devices, was implemented on the Le2i dataset. The model achieved an mAP50 of 0.866 and an mAP50-95 of 0.555, with only 2.52 M parameters and 6.0 GFLOPs. These provide a strong foundation for further improvement in elderly fall detection systems