Human Activity Detection using Wi-Fi RSSI

Introduction

Objective

The goal of this project is to recognize human activities — specifically detecting an empty room, moving, or stationary individuals — using Wi-Fi Received Signal Strength Indication (RSSI) without the need for wearable devices.

Motivation

Privacy concerns surrounding camera-based monitoring systems drive the need for low-cost, non-invasive, and privacy-friendly solutions. This project leverages Wi-Fi RSSI data and machine learning (ML) techniques to address these needs.

Technology

• Wi-Fi RSSI (Received Signal Strength Indication)
• Machine Learning (Random Forest, Convolutional Neural Network)

Hardware and Software Requirements

Hardware

• Four ESP32 S2-mini microcontrollers
  • 1x acts as an Access Point (AP) emitter
  • 3x act as receivers (R1, R2, R3)

Software

• Python (for data processing and model training)
• TensorFlow/PyTorch (for CNN implementation)
• Scikit-learn (for Random Forest model)

Methodology

1. System Setup

• Environment: Small indoor space
• Hardware configuration: One ESP32 as an AP emitter, three ESP32s as receivers

2. Data Collection

• RSSI Measurement: Gather RSSI values from multiple wireless links between the AP and receivers
• Synchronization: Use Wi-Fi broadcast to trigger simultaneous RSSI collection at the receivers via POST requests
• Observation Window: Define a collection duration (e.g., 10-20 seconds)
• Data Classes:
  • "0" - Empty Room
  • "1" - Moving
  • "2" - Stationary
• Storage: Data saved to an SD card on the ESP32 microcontroller

3. Data Pre-Processing

• Normalization: Standardize RSSI values to zero mean and unit variance
• Labeling: Assign numeric labels to data (0 for empty, 1 for moving, 2 for stationary)

4. Machine Learning Models

• Models:
  • Random Forest
  • Convolutional Neural Network (CNN)
• Training: Split data into 80% training and 20% testing
• Prediction: Classify room state as "empty," "moving," or "stationary"

Experiments and Results

Random Forest Model

• Precision: 0.98 (Empty), 0.96 (Stationary), 0.98 (Moving)
• Recall: 0.98 (Empty), 0.98 (Stationary), 0.96 (Moving)
• F1-Score: High consistency across classes
• Support: 126 (Empty), 147 (Stationary), 164 (Moving)
• Overall Accuracy: 97%
• Macro and Weighted Averages: Precision, recall, F1-score ~0.97-0.98

Convolutional Neural Network (CNN)

• Precision: 0.99 (Empty), 0.97 (Stationary), 0.99 (Moving)
• Recall: 0.99 (Empty), 0.98 (Stationary), 0.98 (Moving)
• F1-Score: Consistently high
• Support: 126 (Empty), 147 (Stationary), 164 (Moving)
• Overall Accuracy: 98%
• Macro and Weighted Averages: Precision, recall, F1-score ~0.98

Comparison

• CNN outperformed Random Forest slightly in all metrics
• Both models proved effective, with CNN excelling in precision and recall for "Empty" and "Moving" classes

Challenges and Lessons Learned

• Position Prediction: We attempted a room grid approach (40cm x 40cm) to predict a person’s exact position, but accuracy was too low due to RSSI's sensitivity to noise, interference, and environmental factors like obstacles and multipath propagation.
• Key Limitation: RSSI alone proved insufficient for reliable localization.

Future Improvements

• Incorporate additional data sources (e.g., time-of-flight, angle of arrival, inertial sensors)
• Integrate multiple sensor modalities for improved accuracy

Conclusion

This project successfully implemented a device-free human activity detection system using Wi-Fi RSSI and machine learning. The system effectively detects room states — empty, moving, and stationary — with high accuracy. Future work involves exploring additional data sources to enhance room positioning accuracy and robustness.