Action Recognition Using Indoor Radar Systems

Yang, S. , Le Kernec, J. and Fioranelli, F. (2019) Action Recognition Using Indoor Radar Systems. IET Human Motion Analysis for Healthcare Applications, London, UK, 26 Jun 2019.

Full text not currently available from Enlighten.

Abstract

Activities of Daily Living (ADL) is essential part of elderly care not only in the event of detecting fall, but also for evaluating the pattern of life of an individual, such as food intake and personal hygiene. Continuous monitoring in healthcare are starting to use mobile wireless technologies, optical cameras and biometric sensors. However, those methods suffer from an integration problem where a lot parts has to come together before becoming a practical solution. Furthermore they require technological breakthroughs before implementation. On-the-shelf indoor radar system is a great alternative without legal issues regarding image rights [1][2]. In this work the radar sensing as one of healthcare modality is explored. The most common method to classify activities using radar is based on micro-Doppler radar signatures. The relative motion of structural components of an object/body generates unique patterns in the time-frequency domain of the radar returns. Therefore, different activities are generating unique distinctive features in micro-Doppler signatures that can be used for classification[3]. A limitation of this approach is that radar data for training and testing are typically captured as an image corresponding to a fixed time duration, a sort of “snapshot” spectrogram where only one activity is performed. This is unsuitable for analysing long sequences of realistic data, where a continuous flow of activities is performed by people, with more or less significant transitions between them. Therefore, learning the temporal relationship in a sequence of physical actions is important for human activity classification in 7/24 fashion. Although the idea of interpreting radar data as temporal sequences rather than images for classification was mentioned in late 90s, this approach is still seldom explored in the open literature, and even more so for raw radar data. In this work, we propose using Long Short Term Memory (LSTM) units in a recurrent neural network to classify six different activities, expanding from the preliminary results of simpler binary classification in our previous work [4]. Our most recent work studied a binary classification problem (differentiating between 2 actions) treating the data as time-sequences. I&Q raw radar data and range maps were used separately to determine which one would offer greater classification accuracy using Long Short-Term Memory Neural Networks. We achieved 99.56% accuracy from the raw data and 97.58% accuracy from the range maps with 5 participants. Further experiments for multiple activity classification were carried out with promising results. Furthermore this method can be used to provide persuasive feedback to end-users to advise and influence behaviours for safer and better practice, when anomalies in their routine are identified (prevention & assistance). This will transform the current paradigm of technologies from “reactive” to “proactive” monitoring systems.

Item Type:Conference or Workshop Item
Status:Published
Refereed:Yes
Glasgow Author(s) Enlighten ID:Fioranelli, Dr Francesco and Yang, Dr Shufan and Le Kernec, Dr Julien
Authors: Yang, S., Le Kernec, J., and Fioranelli, F.
College/School:College of Science and Engineering > School of Engineering > Systems Power and Energy
Related URLs:

University Staff: Request a correction | Enlighten Editors: Update this record