A hybrid human fall detection method based on modified YOLOv8s and AlphaPose.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Existing fall detection systems struggle with small object accuracy and multi-person scenarios.
• Need for efficient and accurate human fall detection in real-world applications.

Purpose of the Study:

• To develop a hybrid fall detection method (HFDMIA-Pose) addressing limitations in current systems.
• To improve accuracy, reduce model size, and enhance applicability for multi-person fall recognition.

Main Methods:

• Utilized a modified YOLOv8s object detector with SPD-Conv and a small object detection layer, employing BCIOU loss.
• Integrated a hybrid algorithm using human skeletal nodes for improved fall recognition accuracy.
• Developed a multi-person fall detection dataset (MPFDD) for comprehensive testing.

Main Results:

• HFDMIA-Pose demonstrated improved accuracy (4.30%) and F1 score (4.57%) over AlphaPose, with a 37.50% increase in FPS.
• Achieved average improvements of 5.33% in accuracy and 5.51% in F1 score compared to other models.
• Showcased significant performance gains, reduced model size, and faster processing speeds.

Conclusions:

• HFDMIA-Pose offers superior performance, accuracy, and efficiency for human fall detection.
• The method's advantages make it ideal for resource-constrained edge environments and diverse daily scenarios.
• This hybrid approach presents a competitive and effective solution for advanced human fall detection.