Mobile Computing, Applications, and Services. 9th International Conference, MobiCASE 2018, Osaka, Japan, February 28 – March 2, 2018, Proceedings

Research Article

SmokeSense: Online Activity Recognition Framework on Smartwatches

Download
310 downloads
  • @INPROCEEDINGS{10.1007/978-3-319-90740-6_7,
        author={Muhammad Shoaib and Ozlem Incel and Hans Scholten and Paul Havinga},
        title={SmokeSense: Online Activity Recognition Framework on Smartwatches},
        proceedings={Mobile Computing, Applications, and Services. 9th International Conference, MobiCASE 2018,  Osaka, Japan, February 28 -- March 2, 2018, Proceedings},
        proceedings_a={MOBICASE},
        year={2018},
        month={5},
        keywords={},
        doi={10.1007/978-3-319-90740-6_7}
    }
    
  • Muhammad Shoaib
    Ozlem Incel
    Hans Scholten
    Paul Havinga
    Year: 2018
    SmokeSense: Online Activity Recognition Framework on Smartwatches
    MOBICASE
    Springer
    DOI: 10.1007/978-3-319-90740-6_7
Muhammad Shoaib1,*, Ozlem Incel2,*, Hans Scholten1,*, Paul Havinga1,*
  • 1: University of Twente
  • 2: Galatasaray University
*Contact email: m.shoaib@utwente.nl, odincel@gsu.edu.tr, hans.scholten@utwente.nl, p.j.m.havinga@utwente.nl

Abstract

In most cases, human activity recognition (AR) with smartphones and smartwatches has been done offline due to the limited resources of these devices. Initially, these devices were used for logging sensor data which was later on processed in machine learning tools on a desktop or laptop. However, current versions of these devices are more capable of running an activity recognition system. Therefore, in this paper, we present SmokeSense, an online activity recognition (AR) framework developed for both smartphones and smartwatches on Android platform. This framework can log data from various sensors and can run an AR process in real-time locally on these devices. Any classifier or feature can easily be added on demand. As a case study, we evaluate the recognition performance of smoking with four classifiers, four features, and two sensors on a smartwatch. The activity set includes variants of smoking such as smoking while sitting, standing, walking, biking, as well as other similar activities. Our analysis shows that, similar recognition performance can be achieved in an online recognition as in an offline analysis, even if no training data is available for some smoking postures. We also propose a smoking session detection algorithm to count the number of cigarettes smoked and evaluate its performance.