Spatial Awareness: Human Presence Detection using WiFi CSI

Turn your WiFi signals into an invisible motion sensor. A Privacy-Preserving Alternative to Cameras.

📌 Project Overview

This project utilizes Channel State Information (CSI) from standard WiFi signals to detect human presence. Unlike cameras, it preserves privacy and works in the dark. By analyzing the distortions in WiFi waves (Phase and Amplitude), our Machin Learning model can distinguish between an Empty Room and Human Activity with high precision.

Key Features

• High Accuracy: Achieved 99.40% test accuracy using a Random Forest Classifier.
• Real-Time Detection: Live Streamlit Dashboard for instant visualization.
• Robustness: Optimized feature engineering (Variance, Entropy, Doppler) to filter out electrical noise.
• Privacy-First: No video or audio recording; only signal physics.

---

🚀 Quick Start

1. Installation

Ensure you have Python 3.8+ installed.

pip install -r requirements.txt

2. Run the Dashboard

Launch the real-time application:

streamlit run app.py

Upload a .dat file (e.g., csi_a12_2.dat) to see the detection in action.

3. Training the Model (Optional)

The core model logic is located in model_tools/. To retrain:

python model_tools/train_random_forest.py

---

🏗️ Project Structure

.
├── app.py                            # Main Application (Streamlit Dashboard)
├── model_tools/                      # Model Training Scripts (The "Core Logic")
│   └── train_random_forest.py        # Random Forest Trainer (99.4% Accuracy)
├── models/                           # Saved Model Artifacts (.joblib, .json)
├── data/                             # Dataset Store
│   └── raw/WiAR/                     # Original Intel 5300 .dat files
├── requirements.txt                  # Dependencies
└── _archive/                         # (LEGACY) Raw scripts & experiments.
                                      # Not used in final production.

Note: The _archive/ folder contains initial data processing pipelines and experimental code. It is preserved for reference but is not part of the final executable flow.

---

📊 Performance Metrics

Our final Random Forest Model (150 Estimators) achieved the following on independent test data:

Metric	Score	Meaning
Accuracy	99.40%	Overall Correctness
Precision	100.0%	Zero False Positives (Reliable)
Recall	98.05%	Highly Sensitive to Motion
Inference Time	<10ms	Real-Time Capable

---

👥 Authors

• Shivansh (230054)
• Rishabh (230158)

Newton School of Technology — Computer Networks + AI/ML Capstone