A Cloud-Based Mobile Data Analytics Framework: Case Study of Activity Recognition Using Smartphone

Abstract

Unobtrusive gathering of personal or environmental data using a smartphone can provide the basis for intelligent assistive services. Continuous gathering of data will result in huge amounts of data, especially if many users are involved. Ideally, one might want to keep a large amount of this raw data for future (and maybe different) analysis, and also analyse the data to produce a compact model which can be used in the smartphone for real-time analysis of new data. This motivates a cloud computing solution where data from many users can be stored and analysed efficiently, and then the compact results of the analysis can be downloaded and used in the smartphone. This cloud-based approach is demonstrated using a case study of an activity monitoring application which might be used, for example, to monitor the daily activities, such as walking or going upstairs, of an at-risk person living alone. The cloud-based machine learning uses multiple classification methods, and, starting from individual training sets, enhances and builds classification models for each individual. The cloud-based system also builds a universal model based on all users which can be used as the initial classification model for a new user. The classification model produced by the cloud-based system is downloaded to the smartphone, and can be used to produce accurate real-time activity analysis. As more data is gathered and continually uploaded to the cloud, the models are adapted using an unsupervised learning approach to produce enhanced models which are then downloaded on to the smartphone for improved real-time activity analysis. The evaluation results indicate that the proposed approach can robustly identify activities across multiple individuals: using model adaptation the activity recognition achieves over 95% accuracy in a real usage environment.