Human Activity Recognition Based on Dynamic Active Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Activity recognition is crucial for assessing health and functional status.
• Current methods face challenges with expensive data labeling and activity variability.
• Supervised learning with fixed labels is often unsuitable due to diverse human behaviors.

Purpose of the Study:

• To propose a dynamic active learning method for activity recognition.
• To address limitations of high annotation costs and fixed label sets.
• To enable adaptive analysis of daily activities, including novel patterns.

Main Methods:

• Developed a dynamic active learning approach for activity recognition.
• Samples are selected based on uncertainty, diversity, and representativeness.
• The method dynamically identifies new activities beyond predefined labels.

Main Results:

• Significantly boosted activity recognition performance.
• Considerably reduced annotation costs.
• Enabled adaptive analysis and detection of novel activities.

Conclusions:

• The proposed dynamic active learning method effectively enhances activity recognition.
• It offers a cost-effective solution by minimizing annotation requirements.
• The approach supports adaptive analysis and the discovery of previously unknown activity patterns.