Semi-Supervised Adversarial Auto-Encoder to Expedite Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) using inertial sensors faces challenges with domain adaptation.
• Adapting HAR classifiers typically requires labeled data for each new user, device, or sensor location, hindering practical application.
• This domain adaptation issue is a significant barrier to the widespread adoption of HAR technologies.

Purpose of the Study:

• To propose a semi-supervised HAR method that improves reconstruction and generation capabilities.
• To enable adaptation to new users, devices, and sensor positions using unlabeled data.
• To achieve robust HAR classification without modifying pre-trained classifiers or requiring new labeled data.

Main Methods:

• A semi-supervised approach is proposed, leveraging Variational Autoencoders (VAE) decoupled with adversarial learning.
• The method focuses on improving reconstruction and generation to facilitate domain adaptation.
• It operates without altering a pre-trained HAR classifier and utilizes unlabeled data.

Main Results:

• The proposed framework demonstrates competitive improvement compared to state-of-the-art baselines on a public benchmark dataset.
• Empirical results show effective handling of new and unlabeled activity data.
• The SAA (Semi-supervised HAR method) achieved a 5% improvement in classification score over existing HAR platforms.

Conclusions:

• The developed semi-supervised HAR method effectively addresses the domain adaptation challenge.
• It enables robust classifier operation under variations in users, activities, and sensor positions without new labeled data.
• This approach offers a practical solution for widespread HAR application adoption.