Scholarlink

<p>In vision-based elderly fall detection, complex backgrounds, occluded body regions, scale variations, and redundant feature computation can lead to missed detections, false detections, and reduced inference efficiency. This study presents FallLite-YOLO, a lightweight detector built on the YOLOv11n architecture for real-time elderly fall detection. First, C3K2-SimAM, a three-dimensional neuron attention module, is introduced into the proposed network through a parameter-free attention formulation. This module enhances the separability between fall-related features and background clutter, thereby reducing missed and false detections. Second, BiFPN-ECA, a multi-scale feature fusion attention module, is incorporated to perform bidirectional cross-scale fusion and local channel attention calibration. This design improves the model’s ability to capture fall-related features across different image scales. In addition, GhostNetV2 is adopted as a lightweight backbone to reduce redundant feature computation while preserving effective feature representation. Experimental results show that, compared with YOLOv11n, FallLite-YOLO improves mAP@0.5 by 3.07%, while reducing the number of parameters and FLOPs by 31.67% and 12.12%, respectively. FallLite-YOLO also improves FPS by 7.44% on the tested GPU-equipped workstation, indicating better real-time inference efficiency at the workstation level.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards real-time fall detection with lightweight vision-based networks

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Haowei Pan,
  • Yin Zhang,
  • Cao Ke,
  • Fanxing Meng

摘要

In vision-based elderly fall detection, complex backgrounds, occluded body regions, scale variations, and redundant feature computation can lead to missed detections, false detections, and reduced inference efficiency. This study presents FallLite-YOLO, a lightweight detector built on the YOLOv11n architecture for real-time elderly fall detection. First, C3K2-SimAM, a three-dimensional neuron attention module, is introduced into the proposed network through a parameter-free attention formulation. This module enhances the separability between fall-related features and background clutter, thereby reducing missed and false detections. Second, BiFPN-ECA, a multi-scale feature fusion attention module, is incorporated to perform bidirectional cross-scale fusion and local channel attention calibration. This design improves the model’s ability to capture fall-related features across different image scales. In addition, GhostNetV2 is adopted as a lightweight backbone to reduce redundant feature computation while preserving effective feature representation. Experimental results show that, compared with YOLOv11n, FallLite-YOLO improves mAP@0.5 by 3.07%, while reducing the number of parameters and FLOPs by 31.67% and 12.12%, respectively. FallLite-YOLO also improves FPS by 7.44% on the tested GPU-equipped workstation, indicating better real-time inference efficiency at the workstation level.