A Pervasive Middleware for Activity Recognition with Smartphones
Activity Recognition (AR) is an important research topic in pervasive computing. With the rapid increase in the use of pervasive devices, huge sensor data is generated from diverse devices on a daily basis. Analysis of the sensor data is a significant area of research for AR. There are several devices and techniques available for AR, but the increasing number of sensor devices and data demands new approaches for adaptive, lightweight and accurate AR. We propose a new middleware called the Pervasive Middleware for Activity Recognition (PEMAR) to address these problems. We implemented PEMAR on a Big Data platform incorporating machine-learning techniques to make it adaptive and accurate for the AR of sensor data. The middleware is composed of the following: (1) Filtering and Segmentation to detect different activities; (2) A human centered adaptive approach to create accurate personal models, leveraging on the existing impersonal models; (3) An activity library to serve different mobile applications; and (4) Activity Recognition services to accurately perform AR. We evaluated recognition accuracy of PEMAR using a generated dataset (15 activities, 50 subjects) and USC-Human Activity Dataset (12 activities, 14 subjects) and observed a better accuracy for personal trained AR compared to impersonal trained AR. We tested the applicability and adaptivity of PEMAR by using several motion based applications.
Table of Contents
Introduction -- Related work -- Middleware for gesture recognition -- Implementation and applications -- Results and evaluation -- Conclusion and future work