Open Access

ARTICLE

Integrating Attention Mechanisms in YOLOv8 for Improved Fall Detection Performance

Nizar Zaghden¹, Emad Ibrahim², Mukaram Safaldin^2,*, Mahmoud Mejdoub³

1 Higher School of Business of Sfax, University of Sfax, Sfax, 3018, Tunisia
2 National School of Electronics and Telecommunications of Sfax, University of Sfax, Sfax, 3029, Tunisia
3 Faculty of Sciences of Sfax, University of Sfax, Sfax, 3018, Tunisia

* Corresponding Author: Mukaram Safaldin. Email:

Computers, Materials & Continua 2025, 83(1), 1117-1147. https://doi.org/10.32604/cmc.2025.061948

Received 06 December 2024; Accepted 24 February 2025; Issue published 26 March 2025

Abstract

The increasing elderly population has heightened the need for accurate and reliable fall detection systems, as falls can lead to severe health complications. Existing systems often suffer from high false positive and false negative rates due to insufficient training data and suboptimal detection techniques. This study introduces an advanced fall detection model integrating YOLOv8, Faster R-CNN, and Generative Adversarial Networks (GANs) to enhance accuracy and robustness. A modified YOLOv8 architecture serves as the core, utilizing spatial attention mechanisms to improve critical image regions’ detection. Faster R-CNN is employed for fine-grained human posture analysis, while GANs generate synthetic fall scenarios to expand and diversify the training dataset. Experimental evaluations on the DiverseFALL10500 and CAUCAFall datasets demonstrate that the proposed model significantly outperforms state-of-the-art methods. The model achieves a mean Average Precision (mAP) of 0.9507 on DiverseFALL10500 and 0.996 on CAUCAFall, surpassing conventional YOLO and R-CNN-based models. Precision and recall metrics also indicate superior detection performance, with a recall of 0.929 on DiverseFALL10500 and 0.9993 on CAUCAFall, ensuring minimal false negatives. Real-time deployment tests on the Xilinx Kria™ K26 System-on-Module confirm an average inference time of 43ms per frame, making it suitable for real-time monitoring applications. These results establish the proposed R-CNN_GAN_YOLOv8 model as a benchmark in fall detection, offering a reliable and efficient solution for healthcare applications. By integrating attention mechanisms and GAN-based data augmentation, this approach significantly enhances detection accuracy while reducing false alarms, improving safety for elderly individuals and high-risk environments.

---

Keywords

---