Boosting Performance and Wearability of Activity Monitoring Systems with Transfer Learning

Parastoo Alinia

Keeping track of the activity behavior of the individuals provides information that is useful to many emerging applications ranging from healthcare to gaming. Wearable technology has shown significant growth in the Internet-of-Things (IoT) era due to the ability in continuous and high precision activity and behavior monitoring. Prior research has proposed numerous machine learning solutions based on the data from wearable devices to track the activity behavior of the users which address tasks such as low-level activity type estimation (e.g., metabolic equivalent of task computation and activity-level classification), high-level activity type recognition (e.g., activities of daily living recognition, or posture tracking), and step counting. However, the performance of the pre-trained machine learning models in wearable monitoring systems degrades significantly when used in the end-user setting without retraining the model with new data. To address this issue, transfer learning has been introduced as a machine learning technique that reuses the knowledge learned from one setting in a different but related setting to boost the performance of the machine learning models. However, due to the diversity of the sensor devices, there is still a limitation in the adoption of wearable monitoring systems across diverse domains such as sensors with different modalities, different users, and across different body locations. This dissertation proposes three computational models to improve the performance and wearability of the activity monitoring systems. The proposed models address the problem of cross-domain adoption of the sensor-based activity monitoring models when there is a lack of labels in the target setting. The pilot application of proposed models includes MET estimation, lying posture tracking, and physical activity recognition. The first model is an instance transfer learning approach for accurate MET computation that increase the different body locations. Second, we propose a relational-knowledge transfer learning approach that significantly boosts the performance of the underlying activity recognition models across different sensor modalities, locations, and users. Finally, we develop a deep learning-based instance transfer learning technique that improves the lying posture detecting accuracy of the wristband trackers by transferring the knowledge from a sensor placed on the chest or thigh.

Boosting Performance and Wearability of Activity Monitoring Systems with Transfer Learning

Files and links (1)

Abstract

Metrics

Details