CASL: Capturing Activity Semantics Through Location Information for Enhanced Activity Recognition.

Area of Science:

• Digital healthcare
• Human activity recognition
• Machine learning for healthcare

Background:

• Monitoring daily activities with portable tools is crucial for digital healthcare, particularly in elderly care.
• Current methods heavily rely on costly labeled activity data, hindering widespread adoption.
• Developing cost-effective activity recognition models remains a significant challenge.

Purpose of the Study:

• To propose a novel semi-supervised active learning method for activity recognition.
• To reduce the dependency on extensive labeled datasets.
• To enhance the performance of activity recognition models through expert collaboration.

Main Methods:

• A semi-supervised active learning approach integrating expert collaboration.
• Utilizing user trajectory data as the sole input for the model.
• Employing a query strategy and data fusion for sample selection and performance enhancement.

Main Results:

• The proposed method achieves 89.07% accuracy on the adlnormal dataset with 200 activities, outperforming baseline methods.
• Performance is comparable to supervised learning methods (91.77% accuracy).
• Ablation studies confirmed the effectiveness of the query strategy and data fusion components.

Conclusions:

• The developed semi-supervised active learning method offers an effective and robust solution for activity recognition.
• This approach significantly reduces the cost and effort associated with data labeling.
• The method shows strong potential for practical applications in digital healthcare and elderly care monitoring.