Privacy-Safe Classroom Engagement Sensor Project

ISEF Category: Embedded Systems

Subcategory: Internet of Things  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Most classroom cameras can spot a face, but they also collect way more than you need. What if a sensor could estimate attention without ever seeing a student’s face? That is the core idea here. You can build a system that reads motion and posture as heat patterns, then processes everything on the device itself.

What Is It?

This project asks whether a tiny infrared sensor can estimate student engagement without using a camera. The sensor does not see faces or clothing details. Instead, it reads a low-resolution heat map, kind of like a blurry chessboard of temperatures. From that pattern, a model tries to infer posture shifts, motion, and whether someone seems focused or distracted.

Think of it like listening to a crowd from outside a room. You may not know every word, but you can still tell when the room gets quiet, restless, or active. Here, the sensor is the listener, and the heat pattern is the sound. An ESP32-S3 is a small microcontroller that can run simple machine-learning models right on the device, so the data never has to leave the board.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear input, a clear output, and a clear performance goal. You can compare different posture states, measure prediction accuracy, and study whether on-device inference stays fast enough for real use. The project connects to privacy, classroom analytics, and edge AI, which gives it real-world value. You also get to learn sensor design, data labeling, and model evaluation, which are useful skills for future engineering projects.

Research Questions

• How does sensor placement affect the accuracy of engagement classification?
• What is the effect of model size on inference latency and accuracy?
• Does adding motion-based features improve classification more than heat-map features alone?
• To what extent can a low-resolution IR thermopile array distinguish attentive posture from distracted posture?
• Which window length of sensor readings gives the best balance of accuracy and responsiveness?
• What is the effect of different classroom lighting or desk arrangements on model performance?
• How does on-device only inference compare with cloud-based processing for delay and privacy risk?

Basic Materials

• ESP32-S3 development board with Wi-Fi and BLE support.
• Low-resolution infrared thermopile array sensor.
• USB cable for data transfer and power.
• Laptop for model training and logging.
• Breadboard and jumper wires.
• Small tripod or mounting bracket for fixed sensor placement.
• Tape measure for consistent sensor distance.
• Printed posture reference sheet for labeling data.
• Digital notebook or spreadsheet for ground-truth labels.

Advanced Materials

• ESP32-S3 development board with PSRAM.
• Low-resolution IR thermopile array with documented field of view.
• Optional inertial measurement unit for motion comparison.
• 3D-printed sensor mount for repeatable angles.
• Environmental temperature sensor for background compensation.
• MicroSD logging module if local storage is needed.
• Logic analyzer for debugging timing issues.
• University workstation or laptop for training tiny transformer models.
• Bench power supply for stable embedded testing.

Software & Tools

• Arduino IDE: Helps you flash the ESP32-S3, test sensor reads, and debug embedded code.
• Python: Lets you clean data, label samples, train models, and calculate accuracy.
• Jupyter Notebook: Organizes your analysis, plots, and model comparisons in one place.
• TensorFlow Lite for Microcontrollers: Runs tiny models on the ESP32-S3 for on-device inference.
• ImageJ: Can help inspect and compare heat-map frames if you export sensor data as images.

Experiment Steps

1. Define the exact engagement states you will measure, such as attentive, shifting, and inactive.
2. Decide how you will collect ground-truth labels so your model trains on real behavior, not guesses.
3. Map the sensor setup so distance, angle, and placement stay fixed across trials.
4. Build a baseline classifier first, then compare it with the tiny transformer model.
5. Plan the metrics you will report, including accuracy, latency, confusion matrix, and memory use.
6. Design privacy checks that confirm no image, face, or audio data leaves the device.

Common Pitfalls

• Labeling engagement from memory instead of from recorded events, which creates noisy training data.
• Moving the sensor between sessions, which changes the heat pattern enough to break comparability.
• Collecting data from too few people, which makes the model memorize one student’s habits.
• Ignoring room temperature drift, which can shift the baseline heat map and confuse the classifier.
• Testing only overall accuracy, which can hide false positives on distracted or attentive states.

What Makes This Competitive

A stronger version of this project goes beyond a simple demo. You can compare multiple feature sets, test different sensor placements, and report confusion matrices instead of one accuracy score. You can also study privacy by showing that the model works with no camera, no cloud upload, and no identifiable data. That kind of careful design makes the project feel like real engineering, not just a gadget.

Project Variations

• Compare engagement detection in seated classroom posture versus small-group discussion posture.
• Replace the tiny transformer with a simpler model and test whether the accuracy drop is worth the speed gain.
• Add room-temperature compensation and measure whether it improves predictions across different classrooms.

Learn More

• NIH PubMed: Search for review articles on privacy-preserving sensing, edge AI, and human activity recognition.
• NASA Open Data Portal: Browse sensor-processing examples and learn how engineers handle noisy real-world measurements.
• MIT OpenCourseWare: Search for embedded systems and machine learning courses to study microcontroller design and tiny models.
• IEEE Xplore: Search abstracts and review papers on infrared thermopile arrays, activity recognition, and on-device inference.
• TensorFlow Lite for Microcontrollers documentation: Learn how small models run on embedded boards and how memory limits shape design.