Fall Detection with the Spatial-Temporal Correlation Encoded by a Sequence-to-Sequence Denoised GAN.

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Falls are a leading cause of injury and death, especially for the elderly, necessitating effective fall alarm systems.
• Existing fall detection systems often fail in real-world scenarios due to lighting variations, camera angles, and privacy concerns.
• Data imbalance in supervised learning poses challenges for fall detection due to the rarity of fall events compared to normal activities.

Purpose of the Study:

• To develop a privacy-preserving, real-world applicable fall detection system for the elderly.
• To address limitations of existing systems by utilizing IR-depth and thermal imaging.
• To overcome data imbalance issues using unsupervised anomaly detection.

Main Methods:

• Utilized IR-depth and thermal images as input, ensuring privacy by not capturing facial details.
• Implemented an unsupervised anomaly detection approach, training models on normal activities only.
• Employed a Generative Adversarial Network (GAN) structure to process sequential frames and predict future frames.
• Incorporated multi-subject detection, motion-triggered real-time detection, occlusion handling, and depth image denoising.

Main Results:

• The system demonstrated state-of-the-art performance in fall detection.
• Successfully addressed practical real-world challenges including lighting invariance and privacy protection.
• Effectively handled scenarios with occluded subjects after a fall.
• The unsupervised anomaly detection approach mitigated data imbalance issues.

Conclusions:

• The proposed fall detection system is effective and robust for real-world aged care applications.
• The use of IR-depth and thermal imaging combined with unsupervised learning offers a promising privacy-preserving solution.
• The system's ability to handle various real-world conditions makes it a significant advancement in elderly fall monitoring.