Investigation of Heterogeneity Sources for Occupational Task Recognition via Transfer Learning

Area of Science:

• Human-Computer Interaction
• Machine Learning
• Occupational Safety

Background:

• Human activity recognition (HAR) is vital for classifying occupational tasks.
• Current HAR methods degrade when training and testing data distributions differ due to real-world heterogeneities.

Purpose of the Study:

• To analyze the impact of four heterogeneity sources (cross-sensor, cross-subject, joint cross-sensor and cross-subject, cross-scenario) on HAR classification performance.
• To evaluate the effectiveness of domain adaptation in mitigating performance degradation caused by these heterogeneities.

Main Methods:

• Simulated electrical line worker tasks using separate and mixed task scenarios.
• Employed a support vector machine (SVM) classifier with domain adaptation.
• Benchmarked domain-adapted SVM against a standard SVM baseline.

Main Results:

• Domain-adapted SVM outperformed the baseline in cross-sensor, joint cross-sensor and cross-subject, and cross-subject scenarios.
• No significant performance improvement was observed for the cross-scenario heterogeneity using domain adaptation.

Conclusions:

• Heterogeneity sources significantly impact HAR classification performance.
• Domain adaptation is a valuable technique for enhancing HAR robustness against sensor and subject variations.
• Further research is needed to address cross-scenario performance limitations.