IEEE - Aerospace and Electronic Systems - December 2022 - 37

StudentHighlight:
DOI. No. 10.1109/MAES.2022.3216262
Benchmarking Classification Algorithms for
Radar-Based Human Activity Recognition
Francesco Fioranelli , Simin Zhu , and Ignacio Roldan , TU Delft, 2628 CD
Delft, The Netherlands
Human Activity Recognition (HAR) with radar sensors
has recently attracted great interest from many researchers.
Initially started with a focus on fall detection, the
scope has been broadened to characterize more in general
patterns of activities of daily living of vulnerable individuals.
This has the objective of looking not only at critical
events, such as indeed falls, but also at more subtle
changes and anomalies in the way each individual walks
and performs daily activities, changes that can be linked
to worsening physical and/or cognitive conditions.
One may ask: Why use radar technology for HAR? An
advantage is the contactless sensing capabilities of radar,
whereby locations and movements of people and their different
body parts can be inferred without them actively
wearing, carrying, or interacting with any sensors. This
can be helpful for users' compliance, especially for users
that might forget to carry sensors with them or recharge
their batteries. Furthermore, radar technology does not
record plain optical images or videos of individuals and
their private environments. This aspect can be beneficial
for the " end-users' perception " of this technology in terms
of privacy.
Despite a significant and still increasing body of
research in radar-based HAR, an important limiting aspect
is the lack of a large, commonly accepted dataset for
benchmarking of algorithms. Each individual research
group often uses their own dataset, collected with proprietary
radar sensors and typically involving a group of volunteers-very
often students, which is inevitably small in
terms of the number of people and number/realism of the
activities collected.
Authors' current address: Francesco Fioranelli, Simin
Zhu, and Ignacio Roldan are with the Department of
Microelectronics, TU Delft, 2628 CD Delft, The Netherlands
(e-mail: F.Fioranelli@tudelft.nl, S.Zhu-2@tudelft.
nl, I.RoldanMontero@tudelft.nl).
Manuscript received 5 July 2022, revised 3 October
2022; accepted 18 October 2022, and ready for
publication 21 October 2022.
Review handled by Federico Lombardi.
0885-8985/22/$26.00 ß 2022 IEEE
DECEMBER 2022
However, the radar community is increasingly tackling
this problem and adopting practices of " open access "
to data, which are very much the standard now in the
image and audio processing communities. Examples of
such practices include the organization of classification
challenges on common datasets at the IEEE and IET radar
conferences (e.g., IET Radar Conference 2020, IEEE
Radar Conference 2022), and the public release of datasets
such as the following.
1) The " Open Radar Initiative " [1] was presented at the
2021 IEEE Radar Conference to promote within the
radar community the sharing of datasets for classification
algorithm benchmarking. Available: https://
doi.org/10.1109/RadarConf2147009.2021.9455239
2) The dataset of " Radar signatures of human activities "
released by the radar researchers in the Communication
Sensing and Imaging (CSI) group at the University of
Glasgow [2], [3] is among the first datasets in radarbased
HAR to be publicly shared and has been downloaded
already more than 1,000 times. Available:
https://researchdata.gla.ac.uk/848/
3) The " DopNet " dataset released by the UCL Radar
Group [4] and containing radar spectrograms of several
human gestures poses an important classification
challenge in the field of human-computer interaction.
Available: https://github.com/UCLRadarGroup/DopNet
4) The " PARrad " dataset, aimed at recording patients'
activities in hospital settings, was released by Ghent
University [5]. Available: https://sumo.intec.ugent.
be/parrad
5) The data released by the Computational Intelligence
for Radar Laboratory, at the University of Alabama
Tuscaloosa [6] provides several datasets. Available:
https://github.com/ci4r
6) The " OPERAnet " dataset, by the University ofBristol
and UCL, contains multimodal data for HAR from
sensors including Wi-Fi network interface cards, passive
Wi-Fi implemented on software-defined radio
platforms, UWB radars, and vision plus infrared data
IEEE A&E SYSTEMS MAGAZINE
37

https://orcid.org/0000-0001-8254-8093 https://orcid.org/0000-0003-0092-9082 https://orcid.org/0000-0003-1862-7963 https://doi.org/10.1109/RadarConf2147009.2021.9455239 https://doi.org/10.1109/RadarConf2147009.2021.9455239 https://researchdata.gla.ac.uk/848/ https://github.com/UCLRadarGroup/DopNet https://sumo.intec.ugent.be/parrad https://sumo.intec.ugent.be/parrad https://github.com/ci4r

IEEE - Aerospace and Electronic Systems - December 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - December 2022