Scholarlink

The falls of the elderly carry serious health implication. Therefore, reliable automatic detection systems are needed. This work introduces an optimized two-stage vision-based framework for fall detection that integrates Faster R-CNN for accurate detection of humans, alongside YOLOv5 and YOLOv10 for efficient fall classification with a high-accuracy. Through extraction of human-centric frames and removing the background noise, our approach exhibits superior performance in comparison to single-stage and sensor-based approaches. Experimental results show that YOLOv10 achieves close to perfect detection, training results for which are 0.934 precision, 0.922 recall, and 0.982 for all classes, while the validation confirms the robustness of results (0.935 precision, 0.921 Remarkably enough, the “Fall” class demonstrates 0.937 precision and 0.936 recall in the training with powerful validation scores (0.938 precision, 0.935 recall), keeping the false alarms to a minimum. The “No Fall” class has good reliability (precision 0.931–0.932, recall 0.907), and balances the performance. Compared to the previous studies, our approach has an advantage of confining human subjects only while increasing applicability in real-world to healthcare settings. The efficiency of the system, confirmed by the capability of real-time processing, can justify its edge deployment in assisted living settings. This work takes fall detection to the new level, combining the robust accuracy with pragmatic deployability and becoming a new standard for vision-based monitoring systems.

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

Design and Implementation of a Hybrid Fall Detection Model: Combining Faster R-CNN Inception V2 with YOLO Object Detection Algorithms in Surveillance Systems

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Benedict Ibe,
  • Dagogo Godwin Orifama,
  • Ali Dan,
  • Ikechukwu Nwagbo Enumah,
  • Dominic Chinedu Ogbuagu,
  • Gbubemi Erics

摘要

The falls of the elderly carry serious health implication. Therefore, reliable automatic detection systems are needed. This work introduces an optimized two-stage vision-based framework for fall detection that integrates Faster R-CNN for accurate detection of humans, alongside YOLOv5 and YOLOv10 for efficient fall classification with a high-accuracy. Through extraction of human-centric frames and removing the background noise, our approach exhibits superior performance in comparison to single-stage and sensor-based approaches. Experimental results show that YOLOv10 achieves close to perfect detection, training results for which are 0.934 precision, 0.922 recall, and 0.982 for all classes, while the validation confirms the robustness of results (0.935 precision, 0.921 Remarkably enough, the “Fall” class demonstrates 0.937 precision and 0.936 recall in the training with powerful validation scores (0.938 precision, 0.935 recall), keeping the false alarms to a minimum. The “No Fall” class has good reliability (precision 0.931–0.932, recall 0.907), and balances the performance. Compared to the previous studies, our approach has an advantage of confining human subjects only while increasing applicability in real-world to healthcare settings. The efficiency of the system, confirmed by the capability of real-time processing, can justify its edge deployment in assisted living settings. This work takes fall detection to the new level, combining the robust accuracy with pragmatic deployability and becoming a new standard for vision-based monitoring systems.