💻 computer science

Why Commodity WiFi Sensors Fail at Multi-Person Gait Identification: A Systematic Analysis Using ESP32

This paper demonstrates that the failure of commodity ESP32 WiFi sensors to accurately identify multiple people's gaits stems primarily from fundamental hardware limitations in sensing quality and spatial diversity rather than algorithmic shortcomings, thereby challenging the viability of low-cost WiFi CSI as a robust multi-user biometric primitive.

Original authors: Oliver Custance, Saad Khan, Simon Parkinson

Published 2026-06-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Oliver Custance, Saad Khan, Simon Parkinson

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 you are standing in a crowded room, trying to listen to six different people walking around while they all talk at once. You can hear the general noise of footsteps, but trying to pick out exactly who is walking where, and identifying them just by their footsteps, is incredibly difficult.

This is essentially what this paper investigated, but instead of ears, they used WiFi signals (specifically, a cheap chip called the ESP32), and instead of footsteps, they looked at how people's walking patterns disturb the invisible waves of WiFi.

Here is a breakdown of their findings using simple analogies:

The Big Question: Can Cheap WiFi "Hear" Multiple People?

Scientists have known for a while that WiFi can identify one person walking. It's like recognizing a friend's voice in a quiet room. But what happens when 10 people are walking around at the same time in a busy classroom?

The researchers wanted to know: Is the problem that our "listening" algorithms (the software) aren't smart enough, or is the problem that the cheap WiFi hardware simply can't hear enough detail to begin with?

The Experiment: The "WiFi Cocktail Party"

The team set up a "cocktail party" scenario using cheap, off-the-shelf WiFi sensors (ESP32 chips). They had up to 10 people walking in two different rooms:

A quiet lab: A controlled space with few distractions.
A real classroom: A messy space with furniture and other WiFi signals bouncing around.

They tried six different mathematical "filters" (algorithms) to try to separate the mixed-up signals, hoping to isolate each person's unique walking style (gait). These filters included methods like FastICA, SOBI, and NMF.

The Results: A "Muddy" Signal

The results were a bit disappointing, but very revealing.

The "Muddy Water" Analogy: Imagine trying to separate red and blue food coloring that has been mixed into a single bucket of water. No matter how clever your filter is, if the water is too murky to begin with, you can't perfectly separate the colors.
The Hardware Limit: The cheap WiFi sensors only have 3 antennas and 52 frequency channels. The researchers found this is like trying to listen to a symphony orchestra with only three ears. There simply isn't enough "spatial resolution" (detail) to tell who is who when everyone is moving at once.
The Algorithm Struggle: They tested six different software methods. The best one (called NMF) managed to identify people correctly about 56% of the time. The worst was only about 39%.
- Think of it this way: If you were guessing the answer on a multiple-choice test with 10 options, you'd get 10% right by luck. Getting 56% is better than luck, but it's far from perfect. It's like a coin flip that's slightly weighted, but still unreliable.

Why Did They Fail? (The Three Main Culprits)

The paper introduces three "diagnostic tools" to explain why the systems failed, which are easier to understand with these metaphors:

Intra-Subject Variability (The "Same Person, Different Day" Problem):
- Analogy: You might walk differently today because you are tired, or tomorrow because you are wearing different shoes. The WiFi signal picked up these tiny changes so much that the system thought the same person was actually two different people. The "noise" of the person's own movement was louder than the "signal" of their identity.
Inter-Subject Distinguishability (The "Look-Alike" Problem):
- Analogy: In a crowd, many people walk with a similar rhythm. The WiFi signals from different people overlapped so heavily (97–99% overlap) that the system couldn't tell Person A from Person B. They were all swimming in the same "soup" of data.
Environmental Sensitivity (The "Room Change" Problem):
- Analogy: A filter that worked perfectly in the quiet Lab failed miserably in the noisy Classroom. The way WiFi bounces off walls and furniture (multipath) changed the signal so drastically that a system trained in one room couldn't recognize the same person in another room.

The Verdict: It's the Hardware, Not Just the Software

The most important conclusion of the paper is this: No amount of clever software can fix a broken microphone.

The researchers found that even the best algorithm (NMF) didn't solve the problem because the hardware constraints were the bottleneck. The cheap sensors simply didn't capture enough detail to separate 10 people walking at once. The "separation" algorithms were just rearranging the muddy water, not actually cleaning it.

What Does This Mean for You?

For Security: If you were hoping to use cheap WiFi routers to continuously authenticate (verify identity) for security in a busy office or airport, this paper says: Don't count on it yet. The technology isn't reliable enough to distinguish between 10 people walking past a sensor.
For Privacy: On the flip side, this is good news for privacy. It suggests that cheap, off-the-shelf WiFi devices probably cannot secretly track and identify everyone in a crowded room with high accuracy. The "super-spy" capability of low-cost WiFi is limited by the physics of the hardware, not just the software.

In short: Trying to identify multiple people walking with cheap WiFi is like trying to identify individual singers in a choir by listening to a single, low-quality microphone in a noisy room. The hardware just isn't good enough to hear the details, no matter how smart the computer is.

Technical Summary: Why Commodity WiFi Sensors Fail at Multi-Person Gait Identification

Problem Statement
While WiFi Channel State Information (CSI) has demonstrated promise for single-person gait identification in controlled settings, its feasibility for multi-person identification on low-cost, commodity hardware remains unclear. The core challenge is signal separation: when multiple individuals walk simultaneously, their gait signatures interfere and mix within the received CSI, creating a superposition that must be decomposed before individual identification can occur. Existing literature suggests that while complex setups (e.g., modified firmware, specialized antenna arrays) can achieve separation, performance on commodity devices is modest. The central research question addressed by this paper is whether poor multi-person performance stems primarily from algorithmic limitations (insufficient separation models) or from fundamental sensing ceilings inherent to commodity WiFi hardware (insufficient spatial and spectral diversity).

Methodology
The authors conducted a systematic empirical study using commodity ESP32 WiFi sensors operating in the 5 GHz band (IEEE 802.11n). The study intentionally avoided phase-stable CSI from specialized NICs to reflect realistic low-cost deployment constraints.

Data Collection: Experiments were conducted in two environments: a controlled Laboratory (7m × 7m) and a realistic Classroom (5.4m × 9m) with furniture and ambient traffic. Ten participants (ages 20–40) performed walking trials across seven scenarios involving 1 to 10 people.
Signal Configuration: The system captured CSI amplitude from $k=3$ antennas across $m=52$ information-bearing subcarriers.
Separation Algorithms: Six distinct blind source separation (BSS) methods were evaluated under identical acquisition conditions to isolate algorithmic sensitivity from sensing limitations:
1. FastICA (Statistical independence)
2. SOBI (Second-order temporal structure)
3. PCA–ICA (Hybrid dimensionality reduction and separation)
4. NMF (Non-negative Matrix Factorization; additive parts-based representation)
5. Wavelet Transform (Multi-resolution time-frequency decomposition)
6. Tensor Decomposition (Multi-way factorization preserving interactions across modes)
Evaluation Pipeline: Separated sources were processed into feature vectors (temporal, frequency, and spatial descriptors) and classified using a Support Vector Machine (SVM) with an RBF kernel.
Diagnostic Metrics: Beyond standard accuracy, the authors introduced three metrics to characterize failure modes:
- Intra-Subject Variability (ISV): Consistency of features for the same user.
- Inter-Subject Distinguishability (ISD): Separation between different users in feature space.
- Performance Degradation Rate (PDR): Robustness as crowd density increases.

Key Results
The study found that all evaluated methods performed moderately, with no single algorithm solving the underlying problem of dense multi-person separation.

Classification Performance: The best-performing method was NMF, achieving 56.00% accuracy and a 48.00% F1-score. FastICA and SOBI followed with ~48–49% accuracy, while PCA–ICA performed worst at 39.39%.
Feature Overlap: A critical finding was the extremely high feature-space overlap (97–99%) across all methods. This indicates that even the "best" separation algorithms operate in a regime of strong identity entanglement rather than clean separation.
Stability vs. Separability Trade-off:
- Tensor Decomposition achieved the highest Inter-Subject Distinguishability (ISD) but suffered from extreme Intra-Subject Variability (ISV), meaning the representation for a single user was unstable across trials.
- PCA–ICA offered the most stable representations (lowest ISV) but failed to provide sufficient discriminative power for identification.
- NMF provided the most favorable balance, improving classification metrics without the severe instability of tensor models or the low accuracy of PCA.
Environmental Sensitivity: Performance was highly environment-dependent. Most methods improved in the "richer" multipath Classroom environment, whereas Tensor decomposition performed better in the controlled Laboratory, suggesting that multipath diversity acts as either helpful signal or structured interference depending on the algorithm.
Scalability: As the number of people increased from 2 to 10, all methods showed performance degradation. Tensor decomposition showed the best resilience to density (lowest degradation rate), while Wavelet and PCA–ICA degraded more sharply.

Significance and Claims
The paper concludes that the bottleneck for multi-person gait identification on commodity ESP32 devices is fundamentally a sensing limitation, not merely an algorithmic one.

Hardware Constraints: The limited number of antennas (3) and subcarriers (52) on commodity hardware provides insufficient spatial and spectral degrees of freedom to separate dense mixtures of human gait. The problem is underdetermined; the signal does not contain enough usable information to disentangle users reliably.
Algorithmic Limits: Increasing model sophistication (e.g., moving from matrix to tensor decomposition) does not monotonically improve performance because algorithms cannot compensate for missing information in the measurements. Algorithms primarily redistribute identity entanglement rather than resolving it.
Security and Privacy Implications:
- Authentication: Commodity WiFi CSI is currently a weak foundation for robust multi-user biometric authentication. The high overlap and environmental sensitivity make it unreliable for security decisions requiring stable identity separation.
- Privacy: The results place important bounds on privacy threats. While passive RF sensing is a concern, the study suggests that dense, passive identification using low-cost, off-the-shelf WiFi hardware is less practical than often assumed. The sensing signal itself does not appear to support reliable identity separation in crowded, realistic environments.

Conclusion
The authors assert that under commodity ESP32 CSI constraints, dense multi-person gait identification is limited more by sensing quality and spatial diversity than by the choice of separation algorithm. Future progress in this domain likely requires richer sensing modalities (e.g., massive MIMO, mmWave) or sensor fusion, rather than incremental refinements of separation algorithms on existing low-cost hardware.