The use of WiFi signals for occupancy estimation has gained traction as a “zero-infrastructure” approach to understanding space utilization. However, when evaluated through the lens of measurement theory and real-world deployment constraints, WiFi-based counting systems reveal fundamental limitations. This article synthesizes internal experimental data and academic literature to demonstrate that WiFi is best understood as a probabilistic proxy, not a deterministic sensing modality, and therefore unsuitable for applications requiring high-fidelity occupancy data.
Introduction: The Appeal of Passive Sensing
The WiFi-based occupancy estimation discussed in this article relies on the passive detection of probe requests, which are management frames broadcast by devices searching for access points. Because these signals exist everywhere in modern environments, the approach promises scalability without additional hardware.
Academic research has shown that such systems can produce useful approximations of crowd dynamics, particularly in controlled or large-scale environments . However, this promise often rests on implicit assumptions about signal stability, device identity, and user behavior (assumptions that rarely hold in operational settings).
Learnings from Real Field Testings
To move beyond theoretical assumptions, we conducted controlled experiments comparing WiFi-based estimation against a reference flow counting system with 98% accuracy deployed in the same environment. The objective was not just to observe trends, but to quantify accuracy, stability, and reproducibility under real operating conditions.
Over 2 weeks, both systems captured occupancy dynamics in parallel. The results revealed a consistent pattern: WiFi data closely followed the shape of real occupancy variations, but diverged significantly in absolute values.
- Strong trend alignment: correlation ≈ 0.93
- But large absolute error: up to ±57% vs ground truth
- Even after correction: ±27% residual error
This confirms an important point: WiFi is effective at detecting relative changes in activity, but not at measuring actual occupancy levels.
A deeper look at the time-series data provides further insight. The WiFi signal exhibits noticeable oscillations and discontinuities, even during periods where real occupancy remains stable. These fluctuations are directly linked to the intermittent nature of probe emissions, devices temporarily drop out of detection and reappear minutes later, creating artificial variability in the count.
To compensate for systematic bias, we applied a linear regression model calibrated on the observed data. This reduced the average error:
- From ±57% → ±27% relative error
- From ±9.4 → ±3.1 people absolute deviation
While this improves alignment, it introduces a new dependency: the model must be calibrated to a specific environment. In practice, this means performance becomes context-dependent and non-transferable. Changes in user behavior, device distribution, or even firmware updates can invalidate the correction factor.
Overall, the results are consistent across multiple observations:
- WiFi provides a stable directional signal (when occupancy goes up or down)
- But remains biased and noisy in magnitude (how many people are present)
- And requires continuous calibration to maintain acceptable performance
Why Accuracy Breaks Down
Several independent factors contribute to this gap. Each is well understood individually, but their combined effect makes reliable counting extremely difficult.
1. Identity is no longer stable
Modern smartphones use MAC address randomization, meaning a single device may appear as multiple identities (or none at all).
- Leads to both overcounting and undercounting
- Breaks longitudinal tracking
- Requires complex correction models
This is not a temporary limitation, it is a deliberate privacy feature that will only become more pervasive.
2. Signals are intermittent by design
WiFi probe emissions are not continuous. Their frequency depends on device state, OS behavior, and power management.
- Emission intervals range from seconds to several minutes
- Devices can temporarily “disappear” from the dataset
- Creates noise and oscillation in occupancy curves
Filtering techniques can smooth the signal, but at the cost of latency and responsiveness.
3. Devices are not people
The fundamental assumption 1 device = 1 person does not hold.
- Some individuals carry multiple devices such as laptop, tablets, smart watches
- Others carry none or remain undetected
- Visitor behavior varies widely
This introduces a non-stationary bias that cannot be fully corrected with static calibration.
4. Radio signals ignore physical boundaries
WiFi signals propagate through walls and across spaces.
- Detection extends beyond the intended zone
- External devices (e.g., outside the building) can be counted
- Spatial filtering becomes highly sensitive and environment-specific
This makes precise, zone-level occupancy estimation unreliable.
The Calibration Trap
To compensate for these issues, WiFi systems rely on calibration models.
- Example: correction factor ≈ ×2.5 based on regression
- Error improves, but remains significant (~27%)
The deeper issue is that calibration is not transferable.
Any change in environment, device mix, or user behavior invalidates the model, requiring continuous re-tuning.
A More Accurate Framing
From a scientific perspective, WiFi-based occupancy should be understood as:
- A probabilistic proxy, not a direct measurement
- Suitable for: Trend analysis and long-term behavioral insights
But not for:
- Real-time control systems
- Precise Heatmapping
- Energy optimization loops
- Compliance or safety use cases
- Precise space utilization decisions
A Note on RSSI and CSI-based Approaches
It is worth noting that not all WiFi-based occupancy methods rely on simple device detection. More advanced approaches use Received Signal Strength Indicator (RSSI) or Channel State Information (CSI) to infer human presence based on how bodies affect radio wave propagation.
- RSSI-based methods attempt to estimate proximity or movement by analyzing signal attenuation. However, RSSI is highly sensitive to environmental factors such as multipath reflections, obstacles, and interference, making it unstable and difficult to calibrate in dynamic spaces. Often ~30-50% error in real-world conditions cf Schauer et al., (WiFi-based crowd estimation using RSSI), Farrahi Vahid et al., (device-free passive crowd density estimation)
- CSI-based methods, often explored in academic research, provide much richer physical-layer information and can achieve high accuracy in controlled environments. These techniques analyze fine-grained signal distortions caused by human movement and can, in theory, detect presence without relying on device ownership – Academic Papers suggest ~80-90% accuracy in the best case scenario in a controlled environment cf Zhang et al. – WiFi-based human counting using CSI and Wang et al.– deep learning for multi-person detection with CSI
However, both approaches face significant challenges in real-world deployment:
- CSI requires access to low-level WiFi chipset data, which is rarely available on commercial infrastructure
- Models are environment-specific and fragile, requiring recalibration when conditions change
- Performance degrades in complex, multi-user environments with interference and noise
As a result, while RSSI and CSI demonstrate promising results in laboratory settings, their reliability and scalability remain limited in operational building environments.
Conclusion
WiFi-based occupancy estimation remains an elegant concept, and a useful tool in specific contexts. But its limitations are not marginal: they are structural.
It can indicate patterns.
It can suggest trends.
But it cannot reliably answer the most important question:
How many people are actually present?
For applications where accuracy matters, that distinction is no longer optional but it is foundational.
The Author: Baptiste Potier is the Product Director at Terabee, leading the strategy and development of the product range. He holds an engineering degree from a French engineering school and a Master of Science in Embedded Systems and Robotics, providing a strong technical foundation. His background includes hands-on experience in application development, computer vision, and software engineering.
The Author: Dr. Max Ruffo is a visionary technology leader with over two decades of experience at the forefront of industrial innovation, having pioneered the introduction of 3D printing, civil drones, autonomous mobile robots and LiDAR sensors. Today, Max is dedicated to a long-term mission of building a better world by championing green buildings and net-zero communities.
References and Links
If the topic is of interest, I advise diving into some interesting research with these key leaders:
Academic Research on Waste and Behavior
- (1) Dr. Shahzeen Attari’s research: Her work at Indiana University focuses on people’s judgments and decisions about climate change and resource use. You can explore her publications and profile at:
https://oneill.indiana.edu/faculty-research/directory/profiles/faculty/full-time/attari-shahzeen.html.
Psychological Concepts (Inattentional and Change Blindness)
- (5) Elizabeth Loftus’s work: Her research on memory, inattentional blindness, and change blindness is foundational to this topic.
- „Planting misinformation in the human mind: A 30-year investigation of the malleability of memory“ (2005): https://pubmed.ncbi.nlm.nih.gov/16027179/
- „Change blindness and eyewitness testimony“ (2010):
https://psycnet.apa.org/record/2010-18400-006
Industry and Policy Resources
- (2) American Council for an Energy-Efficient Economy (ACEEE): This non-profit organization provides extensive technical and policy analyses on energy efficiency in buildings. https://www.aceee.org/
- (3) Air Conditioning Contractors of America (ACCA): This association provides resources and standards for the HVAC industry. https://www.acca.org/
- (4) National Comfort Institute (NCI): NCI offers training and resources to HVAC professionals, with a focus on high-performance and energy-efficient systems. https://www.nationalcomfortinstitute.com/
- (6) Commercial Buildings Energy Consumption Survey (CBECS) by the U.S. Energy Information Administration (EIA).
https://www.eia.gov/consumption/commercial/data/2018/ - (7) Eurostat article „Final energy consumption in services – detailed statistics“ https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Final_energy_consumption_in_services_-_detailed_statistics
