8th International Conference on Body Area Networks

Research Article

Cost-Effective Activity Recognition on Mobile Devices

  • @INPROCEEDINGS{10.4108/icst.bodynets.2013.253656,
        author={bin xu and Jian Cui},
        title={Cost-Effective Activity Recognition on Mobile Devices},
        proceedings={8th International Conference on Body Area Networks},
        publisher={ICST},
        proceedings_a={BODYNETS},
        year={2013},
        month={10},
        keywords={activity recognition feature extraction cost reduction},
        doi={10.4108/icst.bodynets.2013.253656}
    }
    
  • bin xu
    Jian Cui
    Year: 2013
    Cost-Effective Activity Recognition on Mobile Devices
    BODYNETS
    ACM
    DOI: 10.4108/icst.bodynets.2013.253656
bin xu, Jian Cui1,*
  • 1: Tsinghua University
*Contact email: cuijian.hh@gmail.com

Abstract

Activity recognition using motion sensors, which denotes a person's posture, has become one of the most important research topics in body sensor network. With the rapid development of monitoring and sensing applications, activity recognition on mobile devices or portable platforms has drawn lots of attentions. Constrained by computing capability and energy budget, activity recognition on mobile devices faces the challenge of hungry energy consumption. Most of the existing work focus on modeling and recognizing activities accurately, however, without computational cost consideration. In this paper, we present WCF: Wavelet Coefficients based Features for activity recognition, which are cost-effective on feature extraction. WCF are extracted from wavelet domain of the sensory raw data and describes the timbre and rhythm properties of activities. Feature space in WCF is hierarchical. Compared with other features, WCF are extracted with less computational cost and redundancy. And WCF have better classification accuracy on activity recognition task. Experiments are conducted on the large public data set USC-HAD to recognize 11 kinds of activities, and our approach outperforms others by reducing 55% ~ 75% computational cost as well as achieving 96.23% classification accuracy.