cs.LGAug 20, 2026
SaveA comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection
Organizations: IDLab, Department of Information Technology, Ghent University-imec, iGent Tower, Technologiepark-Zwijnaarde 126, B-9052 Ghent, Belgium. · Department of Computer Science, Leuven AI, KU Leuven, B-2440 Geel, Belgium · Flanders Make, MPRO, B-3000 Leuven, Belgium · Televic Healthcare, 8870 Izegem, Belgium ·
Abstract
Despite their growing importance for contact-free radio frequency (RF) based healthcare monitoring, different radio technologies such as frequency-modulated continuous wave (FMCW) radar, impulse radio ultra-wideband (IR-UWB), and Wi-Fi sensing are rarely compared under identical deployment conditions, as existing studies typically differ in hardware, datasets, and evaluation methodologies. In addition, the performance of ceiling-mounted radars, despite their practical deployment and cost advantages in healthcare environments, remain underexplored. Therefore, this paper presents a controlled comparison and analysis of ceiling-mounted FMCW, IR-UWB, and Wi-Fi sensing using synchronized recordings from 20 participants across six room layouts. All technologies are evaluated with the same convolutional neural network (CNN) on both a fine-grained 10-class human activity recognition (HAR) task and a coarse 4-class sleep monitoring task. IR-UWB achieves the highest cross-subject activity recognition performance (89.0% macro F1), while FMCW generalizes best to unseen room layouts (83.8% macro F1). For sleep monitoring, all technologies exceed 92% macro F1 in unseen environments. The results reveal a fundamental trade-off between recognition performance and environmental robustness, which can be explained through differences in range resolution, antenna diversity, Doppler resolution, and spatial information retention. These findings provide practical guidelines for the design of healthcare-oriented RF sensing systems.
Figures & tables
| Hardware | Dataset | ||||||||
| Ref. | Year | Type | Device | Setup | #ppl | #rooms | Ceiling mount | #lbl | Nat. flow |
| [ 4 ] | 2023 | IR-UWB | X4M02 | monostatic | 10 | 2 | ✓ | 2 | ✗ |
| [ 5 ] | 2025 | IR-UWB | P440 | monostatic | 10 | 1 | ✓ | 5 | ✗ |
| [ 6 ] | 2024 | Wi-Fi | Intel 5300 | bi-static | 5 | 1 | ✗ | 8 | ✓ |
| [ 7 ] | 2025 | Wi-Fi | Intel 5300 | bi-static | 10 | 2 | ✗ | 6 | ✗ |
| [ 8 , 9 , 10 ] | 2020, 2025 | IR-UWB, FMCW | Xethru X4 Ancortek SDR-Kit TI IWR1443 | monostatic | 6 | \textcolor [gray]0.5 n/r | ✗ | 11 | ✗ |
TABLE I: Comparison of representative RF-based human activity recognition studies in terms of sensing hardware, deployment setup, and dataset characteristics. In contrast to previous work, this study compares FMCW, IR-UWB, and Wi-Fi using synchronized recordings acquired simultaneously in an identical ceiling-mounted setup, enabling a controlled comparison across sensing modalities. ✓/✗ indicate presence/absence of the property. \textcolor [gray]0.5 n/r = not reported. † In this paper, the Wi-Fi transmitter is ceiling-mounted while the receiver is placed on the floor.
Fig. 1: Complete system overview from data acquisition to activity classification. The pipeline illustrates the physical measurement setup (left), the raw signal acquisition and mathematical models for FMCW , IR-UWB , and Wi-Fi (center-left), the modality-specific preprocessing into Time-Doppler and Range-Doppler representations (center-right), and the uniform Convolutional Neural Network (CNN) architecture used to predict the final activity classes.
Fig. 2: Overview of the six room setups used in the study. The bed and chair positions were varied to create different subject positions relative to the fixed ceiling-mounted radar setup. The radar position remained unchanged across all scenarios and is indicated by the coloured markers.
| Parameter | FMCW | IR-UWB | Wi-Fi |
| Antenna setup | |||
| TX antennas | 3 | 1 | 1 |
| RX antennas | 4 | 1 | 1 |
| Signal | |||
| Center frequency | 61.34 GHz | 6.49 GHz | 2.4 GHz |
| Wavelength | 4.94 mm | 46.2 mm | 125 mm |
TABLE II: Configuration and performance parameters of the three sensing modalities.
| Stage | Configuration |
| Input | |
| Conv Block 1 | Conv2D(8, 3 3) + ELU + BN |
| MaxPool(1 2) | |
| Conv Block 2 | Conv2D(16, 3 3) + ELU + BN |
| MaxPool(2 2) | |
| Conv Block 3 | Conv2D(32, 3 3) + ELU + BN |
TABLE III: Uniform CNN architecture applied to all three modalities. The input is a single-channel 2D map of shape , where the two axes correspond to time and Doppler for FMCW and Wi-Fi, and to range and Doppler for IR-UWB. denotes the number of output classes.
| Exp | FMCW | IR-UWB | Wi-Fi | |||||||||
| Fine | Coarse | Fine | Coarse | Fine | Coarse | |||||||
| Before Voting | After Voting | Before Voting | After Voting | Before Voting | After Voting | Before Voting | After Voting | Before Voting | After Voting | Before Voting | After Voting | |
| LOPO | 75.0 9 | 83.4 9 | 88.5 8 | 95.1 3 | 78.6 5 | 89.0 6 | 95.4 3 | 98.2 3 | 65.0 4 | 79.0 7 | 90.2 3 | 96.1 5 |
| LOSO | 77.3 4 | 86.6 1 | 89.9 4 | 94.9 2 | 76.3 5 | 86.4 5 | 95.3 2 | 97.7 3 | 63.2 3 | 75.3 3 | 89.4 2 | 95.1 2 |
| LOBPO | 75.2 2 | 83.8 2 | 88.7 4 | 93.4 2 | 68.3 4 | 78.5 4 | 90.4 5 | 94.2 6 | 57.2 0 | 68.8 0 | 85.5 3 | 92.6 2 |
TABLE IV: Macro F1 scores (%, mean std over folds) for fine-grained activity recognition and coarse-grained sleep monitoring, before and after label voting, under LOPO , LOSO , and LOBPO cross-validation. IR-UWB performs best on the easier cross-subject task, while FMCW is most robust to unseen room layouts. All modalities achieve above 92% macro F1 on the coarse sleep-monitoring task after voting.
| FMCW | IR-UWB | Wi-Fi | ||||||||
| ID | Height | Weight | Age | Sex | Before voting | After voting | Before voting | After voting | Before voting | After voting |
| 1 | 179 | 75 | 30 | M | \textcolor green 88.5 | 90.7 | 82.3 | 89.5 | 69.1 | 91.6 |
| 2 | 189 | 80 | 29 | M | 79.4 | 82.6 | 76.8 | 85.1 | 66.7 | 77.6 |
| 3 | 173 | 82 | 60 | M | 84.7 | 92.3 | 77.9 | 86.3 | 63.2 | \textcolor red 68.7 |
| 4 | 159 | 50 | 40 | F | 71.2 | 82.1 | 72.7 | 87.8 | \textcolor red 52.6 | 72.1 |
| 5 | 182 | 81 | 52 | M | 68.7 | 75.5 | 77.7 | 84.7 | 63.1 | 73.5 |
TABLE V: Per-person leave-one-person-out (LOPO) macro F1 (%) for fine-grained activity recognition before and after label voting, across all three sensing technologies. Demographics: height (cm), weight (kg), age, gender.
Fig. 3: Confusion matrices comparing fine-grained activity classification performance across Wi-Fi, IR-UWB , and FMCW technologies in LOPO and LOBPO cross-validation. Rows indicate true activity labels, columns indicate predicted labels, and values are row-normalised per true class (recall, %).
| Anton Lambrecht received the M.Sc. degree in Computer Science Engineering from Ghent University, Belgium, in 2024. Shortly thereafter, in September 2024, he joined the Department of Information Technology (INTEC) at Ghent University as a researcher within the IDLab research group, where he is currently pursuing the Ph.D. degree. His research interests include RF-based sensing using Ultra-wideband technology, with a focus on healthcare applications through machine learning techniques. |
| Reda el Hail received his Master’s degree in Signal and Image Processing from the University of Rennes 1, France, and his Engineering degree in Electrical Engineering from the Mohammadia School of Engineers in Morocco. As a Ph.D. researcher in machine learning at KU Leuven in Belgium, his current research concentrates on machine learning techniques for indoor human monitoring using millimeter-wave radar technology, with a particular emphasis on enhancing model robustness towards different aspects such as unseen environments, multiple persons and open set recognition. |
| XIANJUN JIAO received his bachelor’s degree in Electrical Engineering from Nankai university in 2001 and Ph.D. in communication and information system from Peking University in 2006. Then, he worked in the industry, such as Nokia, Microsoft and Apple, on wireless systems. In 2016, he joined IDLab, a core research group of imec embedded in Ghent University and University of Antwerp, as a senior researcher. At imec, he works on real-time SDR (Software Defined Radio) platform, such as the openwifi project, open-source Wi-Fi chip design, which is used widely for research in wireless networking, sensing and security. His main interests are designing and implementing high performance/efficient PHY and MAC layers, parallel/heterogeneous computation for wireless communications. |
| PIETER CROMBEZ received the M.S. degree in electronic engineering from KU Leuven University, in 2005, and the Ph.D. degree in electronic engineering from KU Leuven University, in 2009, with his research on low power, reconfigurable transceivers for multistandard/multimode applications. He was a Research Assistant with the ESAT-MICAS Laboratory, KU Leuven University. He joined as the Research and Development Project Lead with Televic Healthcare in 2009, where he was responsible for the wireless research for next generation nurse call products. The main focus was on localization techniques that meet the harsh specifications for the healthcare market. Nowadays, holding the position of Research Lead, he is responsible for the long term research activities within Televic Healthcare with focus on wireless communication, localization and sensing. He has been involved in several national and European funded research projects, both as a participant and a project coordinator. He is a (co)author of several publications. |
| Dominique Schreurs (Fellow, IEEE) received the M.Sc. degree in electronic engineering and the Ph.D. degree from the University of Leuven (KU Leuven), Leuven, Belgium, in 1992 and 1997, respectively. She has been a Visiting Scientist with Agilent Technologies, Santa Rosa, CA, USA, ETH Zürich, Zürich, Switzerland, and the National Institute of Standards and Technology, Boulder, CO, USA. She is currently a Full Professor with KU Leuven, where she is also the Chair of the Leuven ICT (the Leuven Centre on Information and Communication Technology). Her current research interests include the microwave and millimeter-wave characterization and modeling of transistors, nonlinear circuits, and bioliquids, and system design for wireless communications and biomedical applications. Prof. Schreurs served as the President of the IEEE Microwave Theory and Techniques Society from 2018 to 2019. She was an IEEE MTT-S Distinguished Microwave Lecturer. She has also served as the General Chair for the Spring Automatic RF Techniques Group (ARFTG) conferences in 2007, 2012, and 2018, and the President of the ARFTG organization from 2018 to 2019. She currently serves as the TPC Chair for the European Microwave Conference and also the Conference Co-Chair for the IEEE International Microwave Biomedical Conference. She was the Editor-in-Chief of the IEEE Transactions on Microwave Theory and Techniques. |
| PETER KARSMAKERS In 2010 Peter Karsmakers received his Ph.D. at KU Leuven, Department of Electrical Engineering. In 2013 as a post-doc, he co-founded the ADVISE research group at Geel campus together with a few other colleagues. From 2018 he is an Assistant Professor at KU Leuven in the Computer Science Department and since 2022 he is principal investigator at Flanders Make. His research focuses on designing machine learning algorithms that consider application-specific constraints. These can for example relate to the computing platform on which the algorithm will be deployed or background knowledge about the machine learning task. |
| Adnan Shahid (Senior Member, IEEE) received the Ph.D. degree in Information and Communication Engineering from Sejong University, South Korea, in 2015. He is currently an Assistant Professor with the Internet Technology and Data Science Laboratory (IDLab), Ghent University–imec, where he leads the “AI/ML for Wireless” research within the IDLab Intelligent Wireless Networking (iWINe) group. He actively contributes to international standardization efforts, including the IEEE P1900.8 Working Group on Training, Testing, and Evaluating Machine-Learned Spectrum Awareness Models and the ATIS Working Group on Generative AI in Telecommunications. He has participated in several major research initiatives, including the DARPA Spectrum Collaboration Challenge (SC2) and multiple H2020, Horizon Europe, and ESA projects. He has authored over 150 publications in leading conferences and journals. His research interests include wireless physical-layer foundation models, decentralized and federated learning, radio resource management, 5G/6G networks, and non-terrestrial networks (NTN). |
| Eli De Poorter is professor at the IDLab research group from imec & Ghent University (https://idlab.technology). His team performs research on wireless communication technologies such as (indoor) localization solutions, connected healthcare, wireless IoT solutions and machine learning for wireless systems. He performs both fundamental and applied research. He currently has around 350 publications and has a h-index of 50 (google scholar) with around 1400 citations per year. Since 2021, Prof. De Poorter has been included yearly in the Stanford University (Stanford, CA, USA), World’s Top 2% Scientists in the domain of Networking & Telecommunications. |
More in this topic
Human Activity Recognition
Human activity recognition classifies actions and behaviors from wearable sensors, video, radio signals, and other measurements.
See all 228 papers