Machine Learning for Human Activity Recognition Using Non-Intrusive Sensors

  • Yordanka Lazarova Karayaneva

    Student thesis: Doctoral ThesisDoctor of Philosophy


    Human activity recognition with non-intrusive sensors is an emerging topic in the field of computer vision, which has led to applications for supporting the older population. The current studies lack a holistic evaluation of data derived from infrared (IR) sensors including multiple layouts, sensors positions, noise analysis, multi-subject activities, and model generalisation. Micro-Doppler radars are also used extensively for human activity recognition, but the majority of studies fall in the supervised learning category. The very few studies associated with unsupervised human activity recognition with micro-Doppler radars suffer from a number of limitations such as intermediate accuracy and an exploration of a few techniques for feature extraction.

    This thesis explores the use of IR sensors and micro-Doppler radars for human activity recognition for healthcare and eldercare applications. An investigation of such data with a variety of feature extraction and classification methods is achieved by using a number of datasets, comprising of multiple scenarios. Hence, these results address the shortcomings of the previous literature by providing a holistic understanding and evaluation of such data for healthcare purposes. The achieved results for the IR sensors data demonstrate the optimum model for feature extraction and classification, optimum sensor position and layout as well as a novel periodic noise reduction technique. The outcomes of this work bring us a step closer to the potential application of such systems in elderly care homes.

    In terms of micro-Doppler radar data, unsupervised learning is studied considering its importance for unlabelled and poorly labelled projects. Two unsupervised feature extraction techniques are proposed, which are comparable with the existing Convolutional Variational Autoencoder (CVAE) architecture in terms of classification accuracy. The proposed methods provide a reasonable trade-off between computational time and accuracy, which makes them attractive for unsupervised applications.

    Finally, a user acceptance questionnaire is administered in a care home to understand the views and needs of older adults towards sensing technology and physical robots. An extension to the Technology Acceptance Model (TAM) is achieved to cover more characteristics also seen from the perspective of health improvement. Trust and control, as well as physical appearance and functionalities of robots are studied. Two age-based groups are distinguished for statistical analysis with interesting findings.

    To conclude, the techniques presented in this thesis significantly increase the usability of IR sensors and micro-Doppler radars for human activity recognition applications. This brings us a step closer to the application of these systems in care home, where their real world significance will be evaluated.
    Date of AwardApr 2022
    Original languageEnglish
    Awarding Institution
    • Coventry University
    SponsorsCoventry University, Data Driven Research and Innovation (DDRI) PhD grant
    SupervisorSara Sharifzadeh (Supervisor), Vasile Palade (Supervisor) & Yanguo Jing (Supervisor)

    Cite this