Scholarlink

<p>Efficient detection of falls is crucial for ensuring personal safety and minimizing potential complications. For instance, methods relying on computer vision generally demand high-performance equipment for deployment, the cost of which can be prohibitively high. Balancing speed and accuracy in object detection algorithms for low-performance devices poses a significant challenge. To address this problem, we propose a lightweight detector, LEFD, designed to recognize falling behavior in multiple scenarios. Building upon the YOLOv8 framework, we propose a novel Tri-point Interactive Pyramid structure to enhance the effectiveness of multi-stage feature fusion. Additionally, a lightweight detection head is introduced to replace the original detection head of YOLOv8, aiming to balance the model’s detection performance and efficiency. Furthermore, the impact of detrimental gradients on the model is alleviated, and the model’s robustness is improved by introducing a pruning theory based on dependency graphs. Experimental results demonstrate that our method surpasses mainstream object detection algorithms in recognizing falling behavior across multiple scenarios while requiring fewer computational resources.</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

LEFD: lightweight and efficient fall detector for multiple scenarios

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Guangfu Zhong,
  • Guogang Li,
  • Dongping Cao,
  • Wei Cui

摘要

Efficient detection of falls is crucial for ensuring personal safety and minimizing potential complications. For instance, methods relying on computer vision generally demand high-performance equipment for deployment, the cost of which can be prohibitively high. Balancing speed and accuracy in object detection algorithms for low-performance devices poses a significant challenge. To address this problem, we propose a lightweight detector, LEFD, designed to recognize falling behavior in multiple scenarios. Building upon the YOLOv8 framework, we propose a novel Tri-point Interactive Pyramid structure to enhance the effectiveness of multi-stage feature fusion. Additionally, a lightweight detection head is introduced to replace the original detection head of YOLOv8, aiming to balance the model’s detection performance and efficiency. Furthermore, the impact of detrimental gradients on the model is alleviated, and the model’s robustness is improved by introducing a pruning theory based on dependency graphs. Experimental results demonstrate that our method surpasses mainstream object detection algorithms in recognizing falling behavior across multiple scenarios while requiring fewer computational resources.