Passive Indoor People Counting by Bluetooth Signal Deformation Analysis with Deep Learning
Abstract
:1. Introduction
- The passive human sensing application cannot extract from the received message the BLE frequency information, which would be needed to better cope with the variation in RSSI according to the transmission channel. Specifically, the application layer does not provide access to the frequency (that is, the channel) chosen by the Adaptive Frequency Hopping (AFH) mechanism of the BLE protocol to send a message. This constraint is related to security and is not expected to be relaxed in the near future.
- WiFi Channel State Information (CSI) can be directly acquired from certain devices using the WiFi multicarrier encoding mechanism, whereas BLE does not support this type of measurement.
- Acquiring RSSI samples at high rates from BLE-based devices is a challenge. Consequently, collecting the substantial amount of data necessary to train a machine learning model requires a sensibly long "idle" period before the model can be deployed, which may not be feasible for many practical applications [21].
- We tested and compared five different deep learning architectures on a large dataset, in order to select the one that performs best.
- Model fine-tuning. Instead of retraining the model from scratch for each deployment, we collected a large and heterogeneous dataset and used it to pre-train and validate the model. For each new deployment, we then acquire a small dataset from the unknown environment and use it to fine-tune the model. The procedure is quick and accurate, since only about 4000 samples are used.
- High accuracy. The proposed approach can obtain a people count accuracy higher than 99% (according to the accuracy metrics used in the literature) or 96% (according to the more strict metric we adopted) with only a few minutes of training data acquisition in a completely unknown indoor environment, provided that a sufficient number of people can be asked to enter and exit the room during a few minutes interval. The appropriate number of people to involve is equal to, or greater than, the largest number of occupants expected to be counted by the model in the specific environment where it is being fine-tuned.
- High resolution of the people count. If we define the resolution of the people count as the minimum difference between two measurements that can be discriminated by the proposed system, we can consider as optimal a resolution equal to 1. The proposed approach has been evaluated and validated by setting the resolution to 1, that is, to the optimal resolution.
- Interpolation ability. At the cost of a reasonable accuracy degradation, the proposed approach can be easily fine-tuned based only on a suboptimal number of cases, that is, for example, by acquiring samples only with 1-5-9 persons in a room and still being able to count from 0 to 10 occupants. In other words, it can estimate a never-seen-before number of occupants, provided that such a number is not too far from those that it has seen. To the best of our knowledge, this result has never been presented in the literature before.
2. Related Work
3. BLE-Based Passive Human Sensing
3.1. Data Gathering and Pre-Processing
3.2. The Proposed Deep Learning Model
- 1.
- A Dense Neural Network model (DenseNN), composed of a stack of three dense layers.
- 2.
- A Convolutionary Neural Network (CNN), composed of three Convolutionary layers.
- 3.
- A Long Short-Term Memory (LSTM) network, composed of two Long Short-term Memory (LSTM) layers.
- 4.
- A hybrid CNN+LSTM network structured as a sequence of three Convolutionary layers followed by two LSTM layers.
- 5.
- A legacy Transformer model with four heads, four Transformer blocks, and .
4. Experimental Setup
- 1.
- Set the target per-class cardinality (PCC) as the median number of elements across the classes.
- 2.
- Subsample all the classes that have a number of elements larger than PCC, so that all classes have at most PCC elements.
- 3.
- Augment the classes that have fewer than PCC elements.
5. Experimental Results
6. Conclusions and Further Work
Supplementary Materials
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
References
- Yang, J.; Pantazaras, A.; Chaturvedi, K.A.; Chandran, A.K.; Santamouris, M.; Lee, S.E.; Tham, K.W. Comparison of different occupancy counting methods for single system-single zone applications. Energy Build. 2018, 172, 221–234. [Google Scholar] [CrossRef]
- Kouyoumdjieva, S.T.; Danielis, P.; Karlsson, G. Survey of Non-Image-Based Approaches for Counting People. IEEE Commun. Surv. Tutor. 2020, 22, 1305–1336. [Google Scholar] [CrossRef]
- Glavin, P.; Bierman, A.; Schieman, S. Private Eyes, They See Your Every Move: Workplace Surveillance and Worker Well-Being. Soc. Curr. 2024, 11, 327–345. [Google Scholar] [CrossRef] [PubMed]
- Liu, J.; Teng, G.; Hong, F. Human Activity Sensing with Wireless Signals: A Survey. Sensors 2020, 20, 1210. [Google Scholar] [CrossRef]
- Youssef, M.; Mah, M.; Agrawala, A. Challenges: Device-Free Passive Localization for Wireless Environments. In Proceedings of the 13th Annual ACM International Conference on Mobile Computing and Networking (MobiCom ’07), Montréal, QC, Canada, 9–14 September 2007; pp. 222–229. [Google Scholar] [CrossRef]
- Iannizzotto, G.; Milici, M.; Nucita, A.; Lo Bello, L. A Perspective on Passive Human Sensing with Bluetooth. Sensors 2022, 22, 3523. [Google Scholar] [CrossRef]
- Li, W.; Vishwakarma, S.; Tang, C.; Woodbridge, K.; Piechocki, R.J.; Chetty, K. Using RF Transmissions From IoT Devices for Occupancy Detection and Activity Recognition. IEEE Sens. J. 2022, 22, 2484–2495. [Google Scholar] [CrossRef]
- Gebru, K.; Rapelli, M.; Rusca, R.; Casetti, C.; Chiasserini, C.F.; Giaccone, P. Edge-based passive crowd monitoring through WiFi Beacons. Comput. Commun. 2022, 192, 163–170. [Google Scholar] [CrossRef]
- Yousaf, J.; Yakoub, S.; Karkanawi, S.; Hassan, T.; Almajali, E.; Zia, H.; Ghazal, M. Through-the-Wall Human Activity Recognition Using Radar Technologies: A Review. IEEE Open J. Antennas Propag. 2024, 5, 1815–1837. [Google Scholar] [CrossRef]
- Laham, S.A.; Baghi, B.H.; Lajoie, P.Y.; Feriani, A.; Herath, S.; Liu, S.; Dudek, G. Device-Free Human State Estimation using UWB Multi-Static Radios. arXiv 2023, arXiv:2401.05410. [Google Scholar]
- Li, C.; Tanghe, E.; Fontaine, J.; Martens, L.; Romme, J.; Singh, G.; De Poorter, E.; Joseph, W. Multistatic UWB Radar-Based Passive Human Tracking Using COTS Devices. IEEE Antennas Wirel. Propag. Lett. 2022, 21, 695–699. [Google Scholar] [CrossRef]
- Nikookar, H.; Prasad, R. UWB Interference. In Introduction to Ultra Wideband for Wireless Communications; Signals and Communication Technology; Springer: Dordrecht, The Netherlands, 2009; pp. 67–92. [Google Scholar] [CrossRef]
- Chen, H.; Zhang, Y.; Li, W.; Tao, X.; Zhang, P. ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information. IEEE Access 2017, 5, 18066–18074. [Google Scholar] [CrossRef]
- Chen, Z.; Zhang, L.; Jiang, C.; Cao, Z.; Cui, W. WiFi CSI Based Passive Human Activity Recognition Using Attention Based BLSTM. IEEE Trans. Mob. Comput. 2019, 18, 2714–2724. [Google Scholar] [CrossRef]
- Huang, H.; Lin, S. WiDet: Wi-Fi based device-free passive person detection with deep convolutional neural networks. Comput. Commun. 2020, 150, 357–366. [Google Scholar] [CrossRef]
- Chowdary, A.; Bazzi, A.; Chafii, M. On Hybrid Radar Fusion for Integrated Sensing and Communication. IEEE Trans. Wirel. Commun. 2024, 23, 8984–9000. [Google Scholar] [CrossRef]
- Ma, Y.; Zhou, G.; Wang, S. WiFi Sensing with Channel State Information: A Survey. ACM Comput. Surv. 2019, 52, 1–36. [Google Scholar] [CrossRef]
- Soto, J.C.; Galdino, I.; Caballero, E.; Ferreira, V.; Muchaluat-Saade, D.; Albuquerque, C. A survey on vital signs monitoring based on Wi-Fi CSI data. Comput. Commun. 2022, 195, 99–110. [Google Scholar] [CrossRef]
- Qin, Y.; Sigg, S.; Pan, S.; Li, Z. Direction-agnostic gesture recognition system using commercial WiFi devices. Comput. Commun. 2024, 216, 34–44. [Google Scholar] [CrossRef]
- Münch, M.; Huffstadt, K.; Schleif, F. Towards a device-free passive presence detection system with Bluetooth Low Energy beacons. In Proceedings of the 27th European Symposium on Artificial Neural Networks (ESANN 2019), Bruges, Belgium, 24–26 April 2019. [Google Scholar]
- Ramirez, R.; Huang, C.Y.; Liao, C.A.; Lin, P.T.; Lin, H.W.; Liang, S.H. A Practice of BLE RSSI Measurement for Indoor Positioning. Sensors 2021, 21, 5181. [Google Scholar] [CrossRef]
- Billah, M.F.R.M.; Saoda, N.; Gao, J.; Campbell, B. BLE Can See: A Reinforcement Learning Approach for RF-Based Indoor Occupancy Detection. In Proceedings of the 20th International Conference on Information Processing in Sensor Networks (Co-Located with CPS-IoT Week 2021) (IPSN ’21), Nashville, TN, USA, 18–21 May 2021; pp. 132–147. [Google Scholar] [CrossRef]
- Iannizzotto, G.; Lo Bello, L.; Nucita, A. Improving BLE-Based Passive Human Sensing with Deep Learning. Sensors 2023, 23, 2581. [Google Scholar] [CrossRef]
- Münch, M.; Schleif, F.M. Device-Free Passive Human Counting with Bluetooth Low Energy Beacons. In Proceedings of the Advances in Computational Intelligence; Rojas, I., Joya, G., Catala, A., Eds.; Springer: Cham, Switzerland, 2019; pp. 799–810. [Google Scholar] [CrossRef]
- Demrozi, F.; Chiarani, F.; Turetta, C.; Kindt, P.H.; Pravadelli, G. Estimating Indoor Occupancy Through Low-Cost BLE Devices. IEEE Sens. J. 2021, 21, 17053–17063. [Google Scholar] [CrossRef]
- Beato Gutiérrez, M.E.; Sánchez, M.M.; Berjón Gallinas, R.; Fermoso García, A.M. Capacity Control in Indoor Spaces Using Machine Learning Techniques Together with BLE Technology. J. Sens. Actuator Netw. 2021, 10, 35. [Google Scholar] [CrossRef]
- Brockmann, F.; Figura, R.; Handte, M.; Marrón, P.J. RSSI Based Passive Detection of Persons for Waiting Lines Using Bluetooth Low Energy. In Proceedings of the 2018 International Conference on Embedded Wireless Systems and Networks (EWSN ’18), Madrid, Spain, 14–16 February 2018; pp. 102–113. [Google Scholar]
- Feng, C.; Arshad, S.; Liu, Y. MAIS: Multiple Activity Identification System Using Channel State Information of WiFi Signals. In Wireless Algorithms, Systems, and Applications; Ma, L., Khreishah, A., Zhang, Y., Yan, M., Eds.; Springer: Cham, Switzerland, 2017; pp. 419–432. [Google Scholar]
- Adib, F.; Katabi, D. See Through Walls with WiTrack: Device-Free Localization and Tracking Using RF Signals. In Proceedings of the 11th USENIX Symposium on Networked Systems Design and Implementation (NSDI), Seattle, WA, USA, 2–4 April 2014. [Google Scholar]
- Faragher, R.; Harle, R. Location Fingerprinting With Bluetooth Low Energy Beacons. IEEE J. Sel. Areas Commun. 2015, 33, 2418–2428. [Google Scholar] [CrossRef]
- Fink, A.; Beikirch, H.; Voß, M.; Schröder, C. RSSI-based indoor positioning using diversity and Inertial Navigation. In Proceedings of the 2010 International Conference on Indoor Positioning and Indoor Navigation, Zurich, Switzerland, 15–17 September 2010; pp. 1–7. [Google Scholar] [CrossRef]
- Tang, C.; Li, W.; Vishwakarma, S.; Chetty, K.; Julier, S.; Woodbridge, K. Occupancy Detection and People Counting Using WiFi Passive Radar. In Proceedings of the 2020 IEEE Radar Conference (RadarConf20), Florence, Italy, 21–25 September 2020; pp. 1–6. [Google Scholar] [CrossRef]
- Kianoush, S.; Savazzi, S.; Rampa, V.; Nicoli, M. People Counting by Dense WiFi MIMO Networks: Channel Features and Machine Learning Algorithms. Sensors 2019, 19, 3450. [Google Scholar] [CrossRef]
- Huang, J.; Xu, L.; Qian, K.; Wang, J.; Yamanishi, K. Multi-label learning with missing and completely unobserved labels. Data Min. Knowl. Discov. 2021, 35, 1061–1086. [Google Scholar] [CrossRef]
- Zeng, A.; Chen, M.; Zhang, L.; Xu, Q. Are Transformers Effective for Time Series Forecasting? arXiv 2022, arXiv:2205.13504. [Google Scholar] [CrossRef]
- Zhang, G.P. Neural Networks for Time-Series Forecasting. In Handbook of Natural Computing; Rozenberg, G., Bäck, T., Kok, J.N., Eds.; Springer: Berlin/Heidelberg, Germany, 2012; pp. 461–477. [Google Scholar] [CrossRef]
- Zhao, X.; Wang, L.; Zhang, Y.; Han, X.; Deveci, M.; Parmar, M. A review of convolutional neural networks in computer vision. Artif. Intell. Rev. 2024, 57, 99. [Google Scholar] [CrossRef]
- Montesinos López, O.A.; Montesinos López, A.; Crossa, J. Convolutional Neural Networks. In Multivariate Statistical Machine Learning Methods for Genomic Prediction; Springer International Publishing: Cham, Switzerland, 2022; pp. 533–577. [Google Scholar] [CrossRef]
- Van Houdt, G.; Mosquera, C.; Nápoles, G. A review on the long short-term memory model. Artif. Intell. Rev. 2020, 53, 5929–5955. [Google Scholar] [CrossRef]
- Han, C.; Park, H.; Kim, Y.; Gim, G. Hybrid CNN-LSTM Based Time Series Data Prediction Model Study. In Big Data, Cloud Computing, and Data Science Engineering; Lee, R., Ed.; Springer International Publishing: Cham, Switzerland, 2023; pp. 43–54. [Google Scholar] [CrossRef]
- Urbinate, E.; Itano, F.; Del-Moral-Hernandez, E. CNN-LSTM Optimized by Genetic Algorithm in Time Series Forecasting: An Automatic Method to Use Deep Learning. In Artificial Intelligence and Soft Computing; Rutkowski, L., Scherer, R., Korytkowski, M., Pedrycz, W., Tadeusiewicz, R., Zurada, J.M., Eds.; Springer: Cham, Switzerland, 2023; pp. 286–295. [Google Scholar]
- Ahmed, S.; Nielsen, I.E.; Tripathi, A.; Siddiqui, S.; Ramachandran, R.P.; Rasool, G. Transformers in Time-Series Analysis: A Tutorial. Circuits Syst. Signal Process. 2023, 42, 7433–7466. [Google Scholar] [CrossRef]
- Su, L.; Zuo, X.; Li, R.; Wang, X.; Zhao, H.; Huang, B. A systematic review for transformer-based long-term series forecasting. Artif. Intell. Rev. 2025, 58, 80. [Google Scholar] [CrossRef]
- Wang, F.; Zhang, F.; Wu, C.; Wang, B.; Ray Liu, K.J. Passive People Counting using Commodity WiFi. In Proceedings of the 2020 IEEE 6th World Forum on Internet of Things (WF-IoT), New Orleans, LA, USA, 2–16 June 2020; pp. 1–6. [Google Scholar] [CrossRef]
- Nikitin, A.; Iannucci, L.; Kaski, S. TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series. arXiv 2023, arXiv:2305.11567. [Google Scholar]
- Iwana, B.K.; Uchida, S. Time Series Data Augmentation for Neural Networks by Time Warping with a Discriminative Teacher. In Proceedings of the 2020 25th International Conference on Pattern Recognition (ICPR), Milan, Italy, 10–15 January 2021; pp. 3558–3565. [Google Scholar] [CrossRef]
- Hossin, M.; Sulaiman, M.N. A Review on Evaluation Metrics for Data Classification Evaluations. Int. J. Data Min. Knowl. Manag. Process 2015, 5, 1–11. [Google Scholar] [CrossRef]
- Kazaj, P.M.; Baj, G.; Salimi, Y.; Stark, A.W.; Valenzuela, W.; Siontis, G.C.; Zaidi, H.; Reyes, M.; Graeni, C.; Shiri, I. From Claims to Evidence: A Unified Framework and Critical Analysis of CNN vs. Transformer vs. Mamba in Medical Image Segmentation. arXiv 2025, arXiv:2503.01306. [Google Scholar]
Architecture | Advantages | Disadvantages | References |
---|---|---|---|
DenseNN | Simple architecture, fast training. Low computational cost. | Poor at capturing temporal dependencies. Limited feature extraction from structured data. | [36] |
CNN | Excellent at extracting spatial features. Efficient with structured input like RSSI matrices. | Limited temporal modeling. May require large datasets for generalization. | [37,38] |
LSTM | Effective for modeling sequential dependencies. Captures temporal dynamics in RSSI variations. | Slower training. May overfit on small datasets. | [39] |
CNN+LSTM | Combines spatial and temporal modeling. High accuracy and robustness. Best performance in BLE signal deformation tasks. | Slightly higher computational cost. Requires careful tuning. | [40,41] |
Transformer | Captures long-range dependencies. Scalable and parallelizable. | High computational demand. Needs large datasets to perform well. | [42,43] |
Classification Models with and Without Fine-Tuning | ||||
---|---|---|---|---|
Model | Accuracy, no FT | F1, no FT | Accuracy, with FT | F1, with FT |
DenseNN | 77.36% | 77% | 79.41% | 79% |
CNN | 84.79% | 85% | 93.77% | 94% |
LSTM | 83.37% | 83% | 89.27% | 89% |
CNN+LSTM | 85.53% | 86% | 96.83% | 97% |
Transformer | 78.93% | 79% | 96.36% | 96% |
Regression Models Without Fine-Tuning | |||
---|---|---|---|
Model | ACPR, Th = 0.5 | ACPR, Th = 1 | MAE |
DenseNN | 39.74% | 55.70% | 0.84 |
CNN | 73.90% | 81.21% | 0.32 |
LSTM | 79.83% | 85.26% | 0.27 |
CNN+LSTM | 81.25% | 85.65% | 0.26 |
Transformer | 47.92% | 61.16% | 0.66 |
Regression Models with Fine-Tuning | ||||
---|---|---|---|---|
Model | Samples per Class | ACPR Th = 0.5 | ACPR Th = 1 | MAE |
DenseNN | 100 | 40.84% | 64.85% | 0.82 |
200 | 41.04% | 69.10% | 0.77 | |
350 | 41.17% | 69.90% | 0.76 | |
500 | 42.03% | 70.01% | 0.75 | |
700 | 42.88% | 71.01% | 0.73 | |
CNN | 100 | 64.24% | 84.85% | 0.49 |
200 | 74.24% | 89.70% | 0.32 | |
350 | 96.60% | 98.10% | 0.10 | |
500 | 96.63% | 98.02% | 0.11 | |
700 | 96.62% | 98.30% | 0.08 | |
LSTM | 100 | 75.15% | 93.33% | 0.30 |
200 | 86.97% | 95.15% | 0.15 | |
350 | 90.83% | 93.75% | 0.1 | |
500 | 92.12% | 98.27% | 0.09 | |
700 | 92.13% | 98.35% | 0.10 | |
CNN+LSTM | 100 | 76.36% | 95.15% | 0.26 |
200 | 86.67% | 96.67% | 0.15 | |
350 | 96.83% | 99.30% | 0.04 | |
500 | 96.63% | 99.28% | 0.05 | |
700 | 96.62% | 99.74% | 0.04 | |
Transformer | 100 | 45.15% | 65.85% | 0.77 |
200 | 49.77% | 70.01% | 0.73 | |
350 | 50.17% | 72.84% | 0.69 | |
500 | 51.15% | 72.78% | 0.68 | |
700 | 51.09% | 72.33% | 0.70 |
Regression Models Fine-Tuned on 3 Classes | ||
---|---|---|
Model | ACPR, Th = 1 | MAE |
CNN | 72.84% | 1.09 |
LSTM | 71.45% | 1.35 |
CNN+LSTM | 79.88% | 0.97 |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Iannizzotto, G.; Lo Bello, L.; Nucita, A. Passive Indoor People Counting by Bluetooth Signal Deformation Analysis with Deep Learning. Appl. Sci. 2025, 15, 6142. https://doi.org/10.3390/app15116142
Iannizzotto G, Lo Bello L, Nucita A. Passive Indoor People Counting by Bluetooth Signal Deformation Analysis with Deep Learning. Applied Sciences. 2025; 15(11):6142. https://doi.org/10.3390/app15116142
Chicago/Turabian StyleIannizzotto, Giancarlo, Lucia Lo Bello, and Andrea Nucita. 2025. "Passive Indoor People Counting by Bluetooth Signal Deformation Analysis with Deep Learning" Applied Sciences 15, no. 11: 6142. https://doi.org/10.3390/app15116142
APA StyleIannizzotto, G., Lo Bello, L., & Nucita, A. (2025). Passive Indoor People Counting by Bluetooth Signal Deformation Analysis with Deep Learning. Applied Sciences, 15(11), 6142. https://doi.org/10.3390/app15116142