Robust Self-Adaptation Fall-Detection System Based on Camera Height.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Existing vision-based fall detection methods lack practicality due to inconsistent camera heights and limited activity recognition.
• Few studies address the impact of varying sensor placement on fall detection accuracy.
• Current methods often rely on unreliable human segmentation, hindering real-world application.

Purpose of the Study:

• To investigate the effect of camera height on fall detection accuracy.
• To develop a robust fall detection system that accounts for variations in sensor placement.
• To improve the reliability and practicality of vision-based fall detection systems.

Main Methods:

• Construction of a diverse dataset including daily activities and fall events from eight participants at five different camera heights.
• Development of Enhanced Tracking and Denoising Alex-Net (ETDA-Net) to enhance tracking and denoising performance.
• Classification of fall and non-fall events using the proposed ETDA-Net model.

Main Results:

• Fall detection accuracy is significantly influenced by camera height.
• ETDA-Net demonstrates robustness and superior performance compared to traditional deep learning methods.
• The proposed method effectively improves tracking and denoising, leading to more reliable fall event classification.

Conclusions:

• ETDA-Net offers a practical and robust solution for vision-based fall detection, adaptable to varying camera heights.
• The study highlights the importance of considering sensor height variations in fall detection system design.
• This research contributes to the development of more reliable automated fall monitoring systems.