Overcoming Data Scarcity in Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Wearable sensors are increasingly prevalent, driving interest in machine learning for Human Activity Recognition (HAR) in healthcare.
• Current HAR methods often struggle with the need for extensive labeled real-world sensor data, which is time-consuming and costly to acquire.

Purpose of the Study:

• To introduce a novel approach for generating synthetic sensor data using 3D engines and Generative Adversarial Networks (GANs).
• To evaluate the efficacy of this synthetic data in overcoming data scarcity challenges for HAR.
• To compare the performance of models trained on synthetic data against those trained on real-world data.

Main Methods:

• Generation of synthetic sensor data via 3D simulation environments and GANs.
• Evaluation of synthetic data quality through various analytical methods.
• Comparison of machine learning model performance (specifically deep neural networks) using synthetic versus real-world data on benchmark and self-recorded datasets.

Main Results:

• Synthetic data significantly improved deep neural network performance, yielding an 8.4% to 73% higher F1-score for less complex activities compared to state-of-the-art results.
• The performance enhancement diminished for more complex activities, as observed in a longer-duration nursing activity dataset.

Conclusions:

• Synthetic sensor data generated from diverse sources holds significant potential for addressing data scarcity in Human Activity Recognition.
• This method offers a viable strategy to enhance machine learning model performance in healthcare applications reliant on sensor data.
• Further research is needed to optimize synthetic data generation for complex activities and diverse healthcare scenarios.