End-to-End Versatile Human Activity Recognition with Activity Image Transfer Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning is crucial for Human Activity Recognition (HAR) across different domains.
• Existing HAR models often fail in versatile cross-domain transfers (e.g., cross-sensor, cross-body-part) due to limitations in scope and reliance on handcrafted features.
• Inequality in sample distribution across domains negatively impacts transfer performance.

Purpose of the Study:

• To propose a novel framework for versatile cross-domain activity recognition.
• To enable end-to-end implementation by automatically extracting adaptive features from activity images.
• To enhance knowledge transfer capabilities by addressing limitations of existing methods.

Main Methods:

• Exploiting adaptive features from activity images, eliminating the need for complex handcrafted features.
• Implementing a two-stage adaptation strategy: pretraining for domain transferability and separability, and re-weighting for balancing sample contributions.
• Conducting comprehensive experiments on public HAR datasets (DSADS, OPPORTUNITY, PAMAP2).

Main Results:

• The proposed framework demonstrates effective end-to-end implementation for versatile cross-domain HAR.
• The two-stage adaptation strategy significantly enhances knowledge transfer capabilities.
• Experimental results validate the framework's effectiveness across diverse HAR datasets.

Conclusions:

• The developed framework offers a more versatile and effective solution for cross-domain HAR.
• By utilizing adaptive features and a strategic two-stage adaptation, the approach overcomes limitations of previous methods.
• This work advances the applicability of HAR models in real-world scenarios with varying data sources.