⚡ electrical engineering

WiFi-Based People Counting Using Beam-Steerable Antennas: A Test-bed Study

This paper presents a test-bed study on using beam-steerable antennas and Wi-Fi Channel State Information (CSI) to enable ubiquitous people counting and re-identification, leveraging the advanced sensing capabilities of future standards like Wi-Fi 7.

Original authors: Riccardo Bersan, Anay Ajit Deshpande, Sanaz Kianoush, Daniele Piazza, Stefano Savazzi

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Riccardo Bersan, Anay Ajit Deshpande, Sanaz Kianoush, Daniele Piazza, Stefano Savazzi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your home's Wi-Fi router isn't just a device that connects your phone to the internet. Imagine it's also a pair of "invisible eyes" that can see how many people are in a room, even if they are behind a wall or in the dark.

This paper describes a test to see if we can use Wi-Fi signals to count people in a house, using a special kind of antenna that can "look" in different directions without moving.

Here is the breakdown of how they did it and what they found, using simple analogies:

1. The "Flashlight" Antenna

Usually, a Wi-Fi router sends signals out in all directions, like a lightbulb in the center of a room. But the researchers used a special router with Beam-Steerable Antennas.

Think of this like a flashlight with a zoom lens. Instead of lighting up the whole room at once, the router can aim its "Wi-Fi beam" specifically at the living room, then switch and aim it at the kitchen.

The Trick: The router doesn't physically move. It just changes the shape of the signal, like a lighthouse beam sweeping across the sea.
The Goal: By aiming these beams at different spots, the system can get a better "snapshot" of where people are standing.

2. How the "Invisible Eyes" See People

When Wi-Fi signals travel through a room, they bounce off walls, furniture, and people.

The Analogy: Imagine shouting in a cave. If the cave is empty, your echo sounds one way. If someone is standing in the middle of the cave, your echo sounds different because your voice hit their body.
The Science: The researchers measured tiny changes in the Wi-Fi signal (called CSI) as people walked around. When a person moves, they block or bounce the signal, creating a unique "fingerprint" on the data. The more people there are, the more complex the signal gets.

3. The Two Ways of Counting (The Brainpower)

The team tested two different ways to process this data to count the people:

Method A: The "Central Brain" (UL-F)
All the raw data from the Wi-Fi devices is sent to the main router (the Access Point). The router acts like a giant super-brain, looking at all the data at once to decide, "Okay, there are 5 people in here."
Method B: The "Teamwork" Approach (DL-SL / Split Learning)
This is like a relay race.
1. The small Wi-Fi devices (like your laptop or phone) do a little bit of the math first. They look at the signal and say, "I think I see 2 or 3 people."
2. They send just their opinion (not the raw data) to the main router.
3. The main router acts as a referee, taking all those opinions and combining them to make the final count.
- Why do this? It keeps the raw data private on the devices and makes the system faster and more scalable.

4. The Test House

They didn't just do this on a computer; they built a real test environment.

The Setting: A large, two-story house (340 square meters) with 11 rooms.
The Experiment: They had people walk, stand, and move in groups. They tested scenarios with 1 person all the way up to 8 people in the same room.
The Result: They recorded how the Wi-Fi signal changed as the crowd grew.

5. What They Found

More Beams = Better Vision: Using just one "flashlight" angle wasn't perfect. But when they used two different beam angles (switching back and forth quickly), the accuracy improved significantly (by about 10%). It's like looking at an object with both eyes instead of one; you get a better sense of depth and detail.
Teamwork Wins: The "Teamwork" approach (Method B) generally worked better than the "Central Brain" approach. It was more accurate at counting high numbers of people (like 8 people moving around).
The Limit: The system works best when people are moving. If everyone stands perfectly still, it's harder to tell them apart, but the system is designed to detect motion.

Summary

The paper proves that you can turn a standard Wi-Fi network into a people-counter. By using smart antennas that can "steer" their signals like flashlights, and by using a smart teamwork system to analyze the data, they successfully counted up to 8 people moving in a large house without using any cameras.

What the paper doesn't say:
The paper does not claim this can identify who the people are (like recognizing your face), nor does it claim it can diagnose health issues. It is strictly about counting how many people are present and detecting that they are moving.

Technical Summary: WiFi-Based People Counting Using Beam-Steerable Antennas

Problem Statement
The paper addresses the challenge of indoor people counting and discrimination using existing Radio Frequency (RF) infrastructure, specifically WiFi networks. While RF sensing (or "WiFi sensing") has shown promise for identifying individuals and tracking movements through Channel State Information (CSI), current systems often rely on static antenna configurations. The authors propose leveraging Pattern-Reconfigurable Antenna Systems (PARS) to dynamically alter radiation patterns. The core problem is to determine the number of co-present human subjects ( $X$ ) in a monitored space and capture their motion patterns by exploiting the variations in CSI caused by body blockage, utilizing beam-steering capabilities to enhance signal coverage and discrimination without requiring hardware modifications to standard client devices.

Methodology
The study was conducted in a 340 sqm, two-story test-house environment (ADANT testhouse) designed to emulate a realistic residential setting with Non-Line-of-Sight (NLOS) conditions between devices.

Hardware Setup: The system consists of a WiFi 7-compliant Access Point (AP) gateway equipped with a smart antenna system supporting beam-steering (specifically two steerable beams per antenna, codes 21 and 42) in the 5GHz and 6GHz bands. The AP is connected to four Station Devices (STAs), each equipped with a single antenna.
Data Collection: The AP and STAs communicate via uplink. The AP collects CSI data ( $H_t(i, k)$ ) representing the phase shift and amplitude attenuation across $S=120$ usable pilot subcarriers. Data was sampled at 250ms intervals (4 samples/sec) over 30-minute sessions. The dataset includes scenarios ranging from an empty room to $X=8$ subjects moving, standing, or walking in groups.
Signal Processing: The authors calculate the variation in CSI amplitude ( $\Delta H_t(i, k)$ ) over a sliding window of 7 consecutive WiFi frames to detect motion and environmental changes.
Machine Learning Architectures: Two distinct processing architectures were proposed and evaluated:
1. UL-F (Uplink Fusion): All CSI data from the four STAs is transmitted to the AP, where a single XGBoost (Extreme Gradient Boosting) decision tree model is trained to classify the number of people.
2. DL-SL (Downlink Split Learning): A split learning approach where local XGBoost models run on each STA to generate preliminary decisions based on local CSI. These partial decisions are transmitted to the AP, where a Convolutional Neural Network (CNN) aggregates them to make a final decision. This approach keeps raw CSI data local to the STAs.

Key Contributions

Beam-Steerable Antenna Integration: The paper demonstrates the integration of PARS technology with unmodified WiFi signals for people counting. It shows that dynamically steering beams (codes 21 and 42) can improve coverage and create distinct RF fingerprints for different occupancy levels.
Split Learning Architecture: The authors propose and validate a "Split Learning" framework (DL-SL) for RF sensing. This architecture distributes the computational load, allowing STAs to perform local inference (XGBoost) while the AP handles aggregation (CNN), addressing scalability and data privacy concerns by keeping raw CSI off the network.
Comprehensive Test-bed Evaluation: The study provides empirical validation within a large-scale, realistic residential environment, testing the system against varying numbers of subjects (1 to 8) and different beam configurations.

Results

Beam Performance: There was a noticeable performance degradation when using beam pattern 21 compared to beam pattern 42. However, the combined use of both beam patterns significantly improved accuracy, increasing it by approximately 10% in scenarios with high subject density (e.g., $X=8$ ).
Architecture Comparison: The DL-SL (Split Learning) architecture consistently outperformed the UL-F (Uplink Fusion) approach. The distributed processing model demonstrated superior accuracy in counting tasks.
Antenna Diversity: Increasing the number of RX antennas at the AP from $M_R=1$ to $M_R=4$ improved overall counting accuracy, as expected.
Confusion Analysis: Confusion matrices revealed that the joint use of both beam patterns was particularly beneficial when discriminating between a high number of targets ( $X > 4$ ), reducing misclassification rates compared to using single beams.

Significance and Claims
The paper claims that Pattern-Reconfigurable Antenna Systems offer a novel opportunity for human-scale passive sensing by channelizing radiation energy to specific areas, thereby improving the robustness of RF sensing in complex environments. The authors assert that their proposed beam-space processing architecture, particularly the split learning model, provides a scalable solution for future RF sensing systems. The study validates that these technologies can support motion detection and people counting services in residential WiFi networks, paving the way for applications in intrusion detection and building automation. The work emphasizes that these capabilities can be achieved using existing WiFi infrastructure with smart antenna upgrades, without requiring changes to the client devices (STAs).