Fall Detection Algorithm Using Enhanced HRNet Combined with YOLO.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Traditional fall detection algorithms struggle with occluded scenes, single-fall judgments, and real-time performance.
• Limitations include insufficient feature extraction and reliance on simplistic detection methods.

Purpose of the Study:

• To develop a robust, top-down fall detection algorithm overcoming limitations of existing methods.
• To improve accuracy and real-time performance, especially in challenging occluded environments.

Main Methods:

• Utilized a lightweight Shufflenetv2 backbone for YOLOv8, incorporating a mixed attention mechanism for enhanced human pose information.
• Integrated BAM-HRNet with channel attention for precise key point extraction.
• Developed a multi-factor discriminant basis including center of mass velocity, trunk-ground angular velocity, and body height-to-width ratio.
• Implemented an automatic voice inquiry mechanism for fall verification.

Main Results:

• Object detection module achieved 64.1% accuracy on COCO and 61.7% on Pascal VOC datasets.
• Key point detection module reached 73.49% accuracy on COCO and 70.11% on OCHuman datasets.
• The proposed fall detection algorithm exceeded 95% accuracy with a frame rate of 18.1 fps on fall detection datasets.

Conclusions:

• The improved YOLOv8 combined with BAM-HRNet significantly enhances fall detection accuracy and real-time performance.
• The multi-factor analysis and voice inquiry mechanism improve the reliability of fall identification, outperforming traditional algorithms.