Cross-Domain Scalable Activity Recognition Models in Smart Environments
Loading...
Links to Files
Permanent Link
Author/Creator
Author/Creator ORCID
Date
2019-01-01
Type of Work
Department
Information Systems
Program
Information Systems
Citation of Original Publication
Rights
Distribution Rights granted to UMBC by the author.
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan thorugh a local library, pending author/copyright holder's permission.
This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan thorugh a local library, pending author/copyright holder's permission.
This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Abstract
The success of Activity Recognition (AR) methodology largely depends on the availability of labeled training samples and adaptability of activity recognition models in cross-domains such as diverse users, heterogeneous devices, and different smart environments. The availability of new era of Internet-of-Things (IoT) devices ranging from smartphones, smartwatches, micro-radars, Amazon Echo in users everyday environments ease the recognition of human activities, behaviors, and occupancy. Nevertheless, the variabilities across emerging sensors, heterogeneities in consumer devices, and inherent variations in users' activities hinder the design and development of scalable activity recognition models. Motivated by this, in this thesis, we investigate the problem of making human activity recognition scalable–i.e., allowing AR classifiers trained in one context to be readily adapted to a different contextual domain. To allow such adaptation without requiring the onerous step of collecting large volumes of labeled training data, we proposed a transfer learning model that is specifically tuned to the properties of convolutional neural networks (CNNs). We designed different variants of this Heterogeneous Deep Convolutional Neural Network (HDCNN) model that help to automatically adapt and learn the model across different domains, such as different users, device-types, device-instances in presence of completely or partially alike activities in source and target. We also extended the above cross-domain activity recognition models to learn the unseen activities using the deep features transfer learning technique while aggregating the domain knowledge from both the source and target domains. Evaluation on real world datasets attested that our proposed cross-domain activity recognition models are able to achieve high accuracy even without any labeled training data in the target domain, and often offer higher accuracy (compared to shallow and deep classifiers) even with a modest amount of labeled training data.