Scholarlink

<p>Falls are a major public health risk for older adults, yet practical fall-detection systems must be accurate, low-cost, and deployable on resource-constrained devices. We propose an event-triggered hybrid pipeline where an abnormal-acceleration threshold activates a camera module, and a two-stage model (YOLOv5 person localization followed by an eight-layer CNN) verifies falls from the cropped region of interest. Experiments use a public dataset with a predefined Train/Val split (374/111), treating Val as a held-out test set; labels (Fall/Walking/Sitting) are mapped to Fall vs. Non-Fall. To prevent leakage, offline augmentation is applied only to the training data (374 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\rightarrow \)</EquationSource> <EquationSource Format="MATHML"><math> <mo stretchy="false">→</mo> </math></EquationSource> </InlineEquation> 1092 effective images), while the held-out test set remains unchanged; hyperparameters are selected using an internal split of the training partition only. On the held-out test set (72 Fall, 39 Non-Fall), the proposed system achieves 80.2% accuracy, 84.7% sensitivity, and 84.7% F1-score. The deployed INT8 model requires 2.3 MB and runs at 14.8 FPS with 125ms end-to-end capture-to-decision latency on an embedded camera platform, enabling timely mobile alerts. The dataset does not provide demographic metadata; thus, stratified analyses are out of scope.</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

Real-time, mobile-compatible, and low-cost fall detection system with deep learning

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Zahra Elmi,
  • Serhat Derya,
  • Alperen Gözüm,
  • Gökhan Güney,
  • Soheila Elmi

摘要

Falls are a major public health risk for older adults, yet practical fall-detection systems must be accurate, low-cost, and deployable on resource-constrained devices. We propose an event-triggered hybrid pipeline where an abnormal-acceleration threshold activates a camera module, and a two-stage model (YOLOv5 person localization followed by an eight-layer CNN) verifies falls from the cropped region of interest. Experiments use a public dataset with a predefined Train/Val split (374/111), treating Val as a held-out test set; labels (Fall/Walking/Sitting) are mapped to Fall vs. Non-Fall. To prevent leakage, offline augmentation is applied only to the training data (374 \(\rightarrow \) → 1092 effective images), while the held-out test set remains unchanged; hyperparameters are selected using an internal split of the training partition only. On the held-out test set (72 Fall, 39 Non-Fall), the proposed system achieves 80.2% accuracy, 84.7% sensitivity, and 84.7% F1-score. The deployed INT8 model requires 2.3 MB and runs at 14.8 FPS with 125ms end-to-end capture-to-decision latency on an embedded camera platform, enabling timely mobile alerts. The dataset does not provide demographic metadata; thus, stratified analyses are out of scope.