Transfer Learning in Human Activity Recognition: A Survey

TL;DR

This survey analyzes transfer learning for sensor-based HAR in smart homes and wearables, addressing data scarcity, sensor heterogeneity, and deployment constraints. It classifies TL methods into instance, feature, parameter, and knowledge-base transfers, and organizes work along heterogeneity, task difference, multi-source, personalization, and source-selection axes. The review reveals uneven exploration across sub-domains, a heavy emphasis on deep learning in wearables, and a need for standardized evaluation, privacy-preserving deployments, and cross-domain foundational frameworks. The authors provide a state-of-the-art synthesis, identify gaps, and propose a roadmap emphasizing self-supervised learning, cross-modality transfer, and deployable systems to advance HAR in real-world settings.

Abstract

Sensor-based human activity recognition (HAR) has been an active research area, owing to its applications in smart environments, assisted living, fitness, healthcare, etc. Recently, deep learning based end-to-end training has resulted in state-of-the-art performance in domains such as computer vision and natural language, where large amounts of annotated data are available. However, large quantities of annotated data are not available for sensor-based HAR. Moreover, the real-world settings on which the HAR is performed differ in terms of sensor modalities, classification tasks, and target users. To address this problem, transfer learning has been employed extensively. In this survey, we focus on these transfer learning methods in the application domains of smart home and wearables-based HAR. In particular, we provide a problem-solution perspective by categorizing and presenting the works in terms of their contributions and the challenges they address. We also present an updated view of the state-of-the-art for both application domains. Based on our analysis of 205 papers, we highlight the gaps in the literature and provide a roadmap for addressing them. This survey provides a reference to the HAR community, by summarizing the existing works and providing a promising research agenda.

Figures (5)

• Figure 1: Concept map, where we present both the scope and the narrative of our survey. The items highlighted are included in the scope. The narrative shows the categorization of the solution space presented in the survey.
• Figure 2: Illustration of the taxonomy of human activities: The interpretation of the term activity varies across different abstraction levels. We plot the terminologies proposed by Bobick bobick1997movement, Nagel nagel1988image, and Vrigkas et al. vrigkas2015review as per their level of abstraction from a granular to abstract direction. We also provide examples to visualize these terminologies.
• Figure 3: Categorizations in Transfer Learning inspired from the works of pan2009survey and weiss2016survey. Here we provide both the problem perspective, where transfer learning approaches are classified based on the characteristics of the problem setting, and the solution perspective, where the classification is based on the kind of knowledge which is transferred across the domains.
• Figure 4: Illustration of the survey methodology: To conduct a comprehensive survey, we make use of literature review tools like Litmaps and Zotero. Litmaps offers functionalities like article seeding, discovery, and visualizations that are particularly helpful for covering all the relevant works in the selected domain, whereas, Zotero offers storage and categorization capacities.
• Figure 5: Literature Map of transfer learning works performing HAR in the application domain of smart homes (top) and wearables (bottom): In this map, the X-axis denotes the year of publication, whereas, the Y-axis denotes the number of citations. This map is helpful for: (i) verifying if the literature review is comprehensive, (ii) filtering out the recent as well as prominent (well-cited) works, and (iii) analyzing the trends in recent past. These maps are generated using Litmap.