From Lab to Real World: Assessing the Effectiveness of Human Activity Recognition and Optimization through...

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Human activity recognition (HAR) algorithms are typically evaluated in controlled environments.
• Real-world HAR faces challenges like noisy, missing sensor data, and natural human behaviors.
• Existing HAR models may not generalize well to diverse, unobserved daily activities.

Purpose of the Study:

• To introduce a real-world HAR open dataset collected from a wristband accelerometer.
• To evaluate the effectiveness of transfer learning for adapting HAR models to new individuals and contexts.
• To highlight the performance gap between lab-trained and real-world HAR models.

Main Methods:

• Collected a novel, unobserved, real-world HAR dataset with participants' autonomous daily activities.
• Trained a convolutional neural network (CNN) on the real-world dataset.
• Applied transfer learning to adapt a general CNN model using limited real-world data.
• Compared performance of models trained on public (MHEALTH) vs. real-world datasets.

Main Results:

• A general CNN model trained on the real-world dataset achieved 80% mean balanced accuracy (MBA).
• Transfer learning improved the MBA to 85% with fewer data.
• A model trained solely on the MHEALTH dataset achieved 100% MBA but dropped to 62% on the real-world dataset.
• Personalizing the MHEALTH-trained model with real-world data improved MBA by 17%.

Conclusions:

• Transfer learning is crucial for adapting HAR models to real-world conditions and diverse users.
• Real-world datasets are essential for robust HAR system development.
• This study demonstrates the potential of transfer learning to bridge the gap between lab-based and real-world HAR performance.