Robust Activity Recognition for Aging Society.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Human activity recognition (HAR) is crucial for Industry 4.0 and elderly healthcare.
• Existing HAR methods require adaptation for unobtrusive elderly monitoring.
• Robustness to environmental interference is a key challenge in real-world applications.

Purpose of the Study:

• To propose a nonintrusive HAR method for long-term, unobtrusive elderly monitoring.
• To develop a method robust to obstructions and nontarget object interference.
• To leverage spatial-temporal information for accurate activity recognition.

Main Methods:

• Skeleton sequence estimation from RGB images.
• Interframe matching algorithm using activity continuity metrics to filter nontarget objects.
• Novel activity encoding based on interframe joint distances.
• Convolutional Neural Network (CNN) for feature learning.
• Data augmentation to prevent overfitting on small datasets.

Main Results:

• The proposed method achieves state-of-the-art performance on public and a noisy activity dataset (NAD).
• Demonstrated robustness to obstructions and nontarget object interference in realistic environments.
• Effective utilization of spatial-temporal information for accurate recognition.

Conclusions:

• The developed nonintrusive HAR method is suitable for elderly healthcare applications.
• The approach offers robustness in realistic environments, utilizing a single ordinary camera.
• This work advances unobtrusive monitoring solutions for eldercare.