πŸ“„ Paper (PDF) πŸ€— HuggingFace ⚑ GitHub

WiFi Sensing-Based Human Activity Recognition
For Smart Home Applications Using Commodity
Access Points

Gad Gad1, Iqra Batool1, Mostafa M. Fouda2, Shikhar Verma3, Zubair Md Fadlullah1

1Department of Computer Science, Western University, London, ON, Canada

2Department of Electrical and Computer Engineering, Idaho State University, Pocatello, ID, USA

3School of Informatics, Kochi University of Technology, Kochi, Japan

4 Datasets
8.8M CSI Packets
100Γ— Faster
52.9% 7-class HAR
96.7% Localization

Abstract

The ubiquitous availability of WiFi-enabled devices in our homes welcomes rapid adoption of WiFi sensing technology for smart home applications. WiFi sensing is a privacy-preserving, resource-aware sensing technology that requires no additional deployment cost and overcomes limitations of traditional sensing modalities. In this study we adopt a two-phase evaluation framework for WiFi sensing across multiple smart home applicationsβ€”room occupancy detection, human activity recognition, and indoor localization. In Phase 1 (feature representation), we compare our proposed rolling-variance transform against raw amplitude and SHARP-based phase sanitisation, showing that rolling variance achieves up to 1.6Γ— higher accuracy without requiring phase information. In Phase 2 (learning model), we compare traditional ML and deep learning architectures, demonstrating that DL modelsβ€”particularly 1D-CNNβ€”best exploit the temporal structure in rolling-variance features.

πŸ€—
Home HAR 7-class activity recognition Β· 2.7M packets
β†’
πŸ€—
Home Occupation 3-class occupancy detection Β· 1.7M packets
β†’
πŸ€—
Office HAR 4-class activity recognition Β· 0.8M packets
β†’
πŸ€—
Office Localization 4-class indoor localization Β· 1.6M packets
β†’

Key Contributions

01
πŸ“Š

Rolling-Variance Transform

A novel time-invariant feature representation extracted from CSI amplitude alone β€” no phase information needed. Achieves up to 1.6Γ— higher accuracy than raw amplitude on challenging HAR tasks.

02
⚑

100Γ— Faster Processing

Rolling variance computes in 0.03s vs. SHARP's 3.5s β€” a 100Γ— speedup β€” while delivering superior accuracy across all evaluated tasks.

03
🧠

Two-Phase Evaluation

Systematic comparison of feature representations (Phase 1) and learning architectures (Phase 2) across 4 datasets and 3 smart-home tasks.

04
πŸ“

Open-Source Datasets & Tools

Four labeled CSI datasets from ESP32-C6 across home and office environments, plus a real-time collection tool β€” all publicly released.

System Architecture

πŸ“‘
ESP32-C6
TX (AP)
s(t)
Indoor Environment
LoS: A₁e-j2Ο€fτ₁
Wall: Aβ‚‚e-j2Ο€fΟ„β‚‚
NLoS: ALe-j2Ο€fΟ„L
🚢
Hm(n)
πŸ“‘
ESP32-C6
RX (STA)

Processing Pipeline

πŸ“¦ CSI
64 sub.
β†’
πŸ’» Host
UART
β†’
πŸ“„ CSV
Storage
β†’
πŸ” Select
52 LLTF
β†’
⏱️ Resample
150 Hz
Hover over a pipeline step to see its visualization

Feature Representations

SHARP Sanitize
Phase β†’ Amplitude
Rolling Variance ⭐
W ∈ {20, 200, 2000}
Raw Amplitude
52 subcarriers
Hover over a feature method to see its transformation

Figure 1: End-to-end system overview showing signal model, data pipeline, and three preprocessing approaches.

Rolling-Variance Feature Extraction

$$\sigma^2_W[n] = \frac{1}{W}\sum_{i=n-W+1}^{n} x[i]^2 - \left(\frac{1}{W}\sum_{i=n-W+1}^{n} x[i]\right)^{2}$$

The rolling variance acts as a high-pass envelope detector: slow baseline drift is suppressed while activity-induced fluctuations are amplified. Computed efficiently via cumulative sums in O(N) time.

Experimental Results

β–Ό Table I: Dataset Overview and Collection Summary
Dataset Env. Task # Cls Split Recorded Packets Duration
Home HAR (train) Home HAR 7 Session holdout Oct 2025 2.7M 232 min
Home HAR (test) Home HAR 7 ~3.5 mo gap Feb 2026 2.7M 233 min
Home Occ. (train) Home Occ. 3 Temporal split Feb 2026 1.1M 100 min
Home Occ. (test) Home Occ. 3 same-session Feb 2026 0.6M 50 min
Office HAR Office HAR 4 % split Oct 2025 0.8M 66 min
Office Loc. (train) Office Loc. 4 File holdout Oct 2025 0.9M 67 min
Office Loc. (test) Office Loc. 4 File holdout Oct 2025 0.7M 57 min
β–Ό Table II: Communication Setup Parameters Comparison
Parameter SHARP Ours
Monitored channel 802.11ac ch. 42 802.11n HT20
OFDM sample duration, T 3.2 Γ— 10⁻⁢ s 3.2 Γ— 10⁻⁢ s
No. OFDM sub-channels, M 256 (245 used) 64 (52 used)
Subcarrier spacing, Ξ”f 312.5 kHz 312.5 kHz
No. monitoring antennas 4 1
β–Ό Table III: Dataset Characterization
Dataset # Cls Silhouette ↑ Fisher ↑
Home HAR 7 βˆ’0.155 0.164
Home Occ. 3 0.412 0.594
Office HAR 4 0.407 4.030
Office Loc. 4 0.800 23.446
Reference Baselines
Iris 3 0.513 6.632
MNIST 10 0.063 0.334
CIFAR-10 10 βˆ’0.058 0.088

Phase 1 Results: Feature Representation

β–Ό Table IV: Full ML Results (Best accuracy per configuration)
Home HAR (7 cls) Home Occ. (3 cls) Office HAR (4 cls) Office Loc. (4 cls)
Pipeline Model L* Acc Pipeline Model L* Acc Pipeline Model L* Acc Pipeline Model L* Acc
AmplitudeRF2k.269 AmplitudeRF5001.00 AmplitudeRF500.815 AmplitudeRF500.906
AmplitudeXGB1k.292 AmplitudeXGB5001.00 AmplitudeXGB2k.833 AmplitudeXGB1k.904
Roll. Var.RF500.463 Roll. Var.RF2k.978 Roll. Var.RF2k.933 Roll. Var.RF500.891
Roll. Var.XGB2k.446 Roll. Var.XGB500.987 Roll. Var.XGB1k.908 Roll. Var.XGB500.868

Phase 2 Results: Deep Learning

β–Ό Table V: Full DL Results (Rolling-variance pipeline)
W Architecture Home HAR Home Occ. Office HAR Office Loc.
L*Acc L*Acc L*Acc L*Acc
20 1D-CNN 1k.529 1k.993 1k.942 2k.957
1D-Conv-LSTM 500.510 2k1.00 500.933 1k.967
MLP 1k.351 500.958 500.836 1k.890
200 1D-CNN 2k.513 2k1.00 500.941 1k.957
1D-Conv-LSTM 2k.525 1k1.00 500.933 500.889
MLP 1k.350 500.989 500.882 1k.924
2000 1D-CNN 500.355 1k.996 500.824 1k.934
1D-Conv-LSTM 500.468 500.994 500.924 500.916
MLP 500.143 500.989 500.483 1k.935

Performance Comparison

ML Accuracy by Feature Pipeline

Home HAR
Home Occ.
Office HAR
Office Loc.
Rolling Variance Amp.+Sanitised Amplitude

ML vs DL Comparison

Home HAR
Home Occ.
Office HAR
Office Loc.
ML Best DL (W=20)

Results Dashboard

All metrics rendered natively from experimental CSVs. Best result per configuration highlighted.

Fig. 6 β€” Reproduced

Feature Pipeline Comparison

Best accuracy (max over window lengths & models) per feature pipeline. Rolling variance outperforms phase-dependent methods on the hardest tasks.

DL Results

DL Architecture Comparison

Best accuracy per DL architecture across all rolling-variance windows and window lengths. Conv1D leads on HAR; CNN-LSTM tops localization.

1D-CNN Architecture

Input (B, 52 Γ— W)
↓
Reshape (B, 52, W)
↓
Conv1D 52β†’64, k=7 β†’ BN β†’ ReLU β†’ Pool
↓
Conv1D 64β†’128, k=7 β†’ BN β†’ ReLU β†’ Pool
↓
Conv1D 128β†’128, k=7 β†’ BN β†’ ReLU β†’ Pool
3Γ— Conv Blocks
↓
Global Average Pool β†’ BN
↓
FC 128β†’64 β†’ ReLU β†’ Dropout
↓
FC 64β†’C β†’ Softmax
Classifier Head

Figure 3: 1D-CNN architecture achieving best overall results on HAR tasks.

Resource Consumption

Table VI: Preprocessing Benchmark

Method Wall (s) CPU (s)
Rolling Var. (W=20) 0.031 0.031
Rolling Var. (W=200) 0.027 0.016
Rolling Var. (W=2000) 0.027 0.031
SHARP Sanitisation 3.508 3.516

Rolling variance is 100Γ— faster than SHARP!

Table VII: Learning Benchmark

Model Train (s) Infer (s) Acc.
RandomForest 20.38 0.36 .449
XGBoost 1795.80 0.34 .476
1D-CNN 185.12 0.89 .454
MLP 245.90 0.39 .382

Ablation Study

Table VIII: Rolling-Variance Window W

W Home HAR Home Occ. Off. HAR Off. Loc.
20 0.529 0.993 0.942 0.957
200 0.512 1.000 0.941 0.957
2000 0.355 0.996 0.824 0.933

Table IX: Window Length L

L Home HAR Home Occ. Off. HAR Off. Loc.
500 0.507 0.984 0.937 0.943
1000 0.529 0.993 0.942 0.953
2000 0.504 0.982 0.917 0.957

Conclusion

🎯

1.6Γ— higher accuracy with rolling variance vs. raw amplitude on challenging HAR tasks

⚑

100Γ— faster than SHARP-based phase sanitisation

🧠

1D-CNN best exploits temporal structure, reaching 52.9% on 7-class Home HAR

πŸ“±

No phase information needed β€” works with amplitude-only CSI

Future work: Cross-environment transfer learning, multi-subject generalization, and on-device deployment on ESP32 microcontrollers.