Transfer Learning of Human Activities Based on IMU Sensors: A Review

This systematic review comprehensively scrutinizes transfer learning (TL) methods applied to human activity recognition (HAR) using inertial measurement unit (IMU) data. Our objective is to provide a comprehensive resource for researchers and developers by summarizing the existing activities, feature extractions, and TL techniques in the related studies. Moreover, we identify research gaps in these categories, allowing research endeavors to address these gaps in the future. Our methodology follows the structure of the preferred reporting items for systematic reviews and meta-analysis (PRISMA) statement by formulating precise research questions, establishing search queries, and specifying inclusion and exclusion criteria for study selection. Finally, we extracted and summarized the existing activities, feature extractions, and TL techniques utilized in the included studies. We analyzed 447 studies from PubMed, ACM, and Scopus datasets, of which we ultimately selected 33 pivotal studies that met our inclusion criteria. Overall, we found that TL has enhanced HAR performance by reusing pretrained models. However, it is important to carefully select relevant transfer information to avoid any potential adverse effects. We conclude that there is a lack of studies assessing specific activities, such as personal hygiene and elder care, using IMU sensors and TL. To improve signal representation, a combination of features based on activity nature is important. Aligning algorithms with activity nature is essential—simpler models like AlexNet are suitable for routine activities such as walking, while more complex models like DenseNet are better for intricate tasks like cleaning. Our review is a reference for advancing the understanding of TL’s potential of HAR using IMU data.