A Semi-Automatic Annotation Approach for Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Modern smartphones and wearables generate vast sensor data for applications like healthcare and indoor location.
• Human Activity Recognition (HAR) algorithms require extensive expert-annotated data, which is costly and time-consuming.
• Developing HAR methods with reduced annotation effort is a significant challenge.

Purpose of the Study:

• To introduce a Semi-Supervised Active Learning (SSAL) approach based on Self-Training (ST) for Human Activity Recognition (HAR).
• To automate parts of the data annotation process, thereby reducing annotation effort and the volume of required labeled data.
• To achieve high-performance HAR classifiers with minimal expert input.

Main Methods:

• Implemented a Semi-Supervised Active Learning (SSAL) strategy incorporating Self-Training (ST).
• Utilized a criterion for selecting the most relevant data samples for expert annotation.
• Propagated expert labels to the most confident samples identified by the model.

Main Results:

• The proposed SSAL approach significantly reduces the need for annotated data.
• Achieved a reduction in required annotated data by more than 89%.
• Maintained accurate model performance comparable to traditional supervised methods despite reduced data.

Conclusions:

• The SSAL approach effectively reduces annotation burden in HAR.
• It is possible to achieve high-performance HAR models with substantially less labeled data.
• This method offers a practical solution for developing HAR systems efficiently.