Magnetic Vision: Teaching AI to See Motion Without a Camera

A TinyML project using a Hall-effect sensor and Edge Impulse to classify motion speed and patterns — no camera required. Runs entirely on-device on a Seeed Wio Terminal.

Table of Contents

1. Overview
2. Hardware Setup
3. Experiment 1: Speed Classification
4. Experiment 2: Motion Pattern Classification
5. Video Demos
6. How to Deploy
7. Repository Structure
8. Results Summary
9. Future Work
10. Credits
11. Links

Overview

This project explores whether a simple Hall-effect sensor can capture enough information about a moving object for a machine-learning model to recognize its motion — without any camera or computer vision.

It contains two experiments:

• Experiment 1: Classify motion speed (FAST / SLOW / NO_CAR) — achieved 100% accuracy using the Flatten processing block.
• Experiment 2: Classify motion patterns (STRAIGHT / OSCILLATE / PAUSE) — achieved 76.3% validation accuracy using a 1D CNN on raw signals.

Key finding: Statistical features (Flatten) work well for amplitude-based problems, but temporal patterns require more sophisticated modeling (1D CNN).

Hardware Setup

Components

Wiring Diagram

Experiment 1: Speed Classification

Classes

• FAST: Car moves rapidly past the sensor.
• SLOW: Car moves slowly past the sensor.
• NO_CAR: No car present near the sensor.

Data Collection

Impulse Design

• Window size: 6000 ms (6 seconds at 100 Hz)
• Window increase: 3000 ms
• Processing block: Flatten (extracts 7 statistical features)
• Learning block: Classification (Keras)

Features Extracted by Flatten

• Mean
• Standard Deviation
• Minimum
• Maximum
• RMS
• Skewness
• Kurtosis

Neural Network

• Input: 7 features
• Dense (20 neurons, ReLU) → Dense (10 neurons, ReLU) → Dropout (0.2)
• Output: 3 classes, Softmax

Results

• Validation accuracy: 100%
• Test accuracy: 100%
• Dataset: 10 recordings per class (30 total) → 90 windows

Deployment

The model was deployed as an Arduino library using the EON Compiler for optimized on-device performance.

• https://studio.edgeimpulse.com/public/1092905/live

Experiment 2: Motion Pattern Classification

Classes

• STRAIGHT: Steady movement in one direction.
• PAUSE: Movement, brief pause, then continuation.
• OSCILLATE: Back-and-forth movement within the window.

Data Collection

Each recording lasts 3 seconds (300 samples at 100 Hz). Below are examples of each class:

The dataset consists of 86 recordings split across three classes:

Processing

• Window size: 3000 ms (3 seconds at 100 Hz)
• Processing block: Raw Data (full 300-sample signal)
• Learning block: Classification (Keras with 1D CNN)

Neural Network

• Input: 300 raw samples → Reshape (300 × 1)
• 1D Conv/Pool (4 filters, kernel 3) → 1D Conv/Pool (16 filters, kernel 3)
• Flatten → Dense (20, ReLU) → Dropout (0.2)
• Output: 3 classes, Softmax

Results

• Validation accuracy: 76.3%
• Test accuracy: 66.67%
• Dataset: 28–29 recordings per class (86 total)

Finding: Flatten was tested first and reached only 61.5%, confirming that amplitude-based statistics are insufficient for temporal pattern recognition.

Deployment

The model was deployed as an Arduino library using the EON Compiler for optimized on-device performance.

• https://studio.edgeimpulse.com/public/1112065/live

On-device performance:

• Inference latency: 5 ms
• Peak RAM usage: 13.8 KB
• Flash usage: 68.5 KB

Video Demos

Experiment 1: Speed Classification (FAST / SLOW / NO_CAR)

• Live demo — take 1
• Live demo — take 2
• FAST detection
• SLOW / NO_CAR detection

Experiment 2: Motion Pattern Classification (STRAIGHT / OSCILLATE / PAUSE)

• Live demo

How to Deploy

Prerequisites

1. Arduino IDE (v1.8.19 recommended for stability)
2. Seeed SAMD board package installed
3. TFT_eSPI library
4. Edge Impulse Arduino library (specific to your impulse)

Step 1: Install the Board

1. Open Arduino IDE.
2. Go to File → Preferences.
3. Add this URL to "Additional Boards Manager URLs":

   https://files.seeedstudio.com/arduino/package_seeeduino_boards_index.json
   

4. Go to Tools → Board → Board Manager.
5. Search for "Seeed SAMD" and install it.
6. Select Tools → Board → Seeeduino Wio Terminal.

Step 2: Install Libraries

1. Install TFT_eSPI via Sketch → Include Library → Manage Libraries.
2. Download the Edge Impulse Arduino library from your project's Deployment page on Edge Impulse Studio.
3. Install it via Sketch → Include Library → Add .ZIP Library...

Step 3: Upload the Firmware

1. Open the desired .ino file from firmware/.
2. Connect the Wio Terminal via USB.
3. Select the correct port in Tools → Port.
4. Press Upload.

Step 4: Run the Project

1. Open Serial Monitor at 9600 baud.
2. Press B1 on the Wio Terminal to start recording.
3. Move the magnet/car during the recording window.
4. The result will appear on the Wio Terminal's screen.

Step 5: PC Logger (Optional)

If you want to collect your own dataset:

1. Install Python 3 and pyserial:

   pip install pyserial
   

2. Edit SERIAL_PORT in pc_tools/logger.py to match your COM port.
3. Run: python pc_tools/logger.py
4. Press B1 on the Wio Terminal; the data will be saved automatically to dataset/<CLASS>/.

Repository Structure

https://github.com/aula9/Magnetic-Vision-Teaching-AI-to-See-Motion-Without-a-Camera

Magnetic-Vision-Teaching-AI-to-See-Motion-Without-a-Camera/
│
├── firmware/              Arduino sketches
│   ├── experiment1_speed/      (FAST / SLOW / NO_CAR)
│   └── experiment2_patterns/   (STRAIGHT / OSCILLATE / PAUSE)
│
├── pc_tools/              Python scripts
│   ├── logger.py               (receives CSV data via Serial)
│   └── visualize.py            (plots time-series signals)
│
├── dataset/               Collected CSV recordings (organized by class)
│
├── images/                Project screenshots and diagrams
│
└── README.md              Project documentation

Results Summary

Future Work

• Multi-axis sensing: Use 3-axis magnetometers to capture richer magnetic signatures.
• Directional detection: Add a second Hall-effect sensor to detect motion direction.
• Larger datasets: Vary magnet size, orientation, speed, and distance to test generalization.

Credits

Author: Aula Jazmati
Platform: Seeed Studio Wio Terminal
ML Pipeline: Edge Impulse
License: MIT

Full Write-up

View the complete project on Hackster