Wi-Fi CSI Activity Recognition Project Ideas

ISEF Category: Embedded Systems

Subcategory: Networking and Data Communications  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A camera can see through a window, and a Wi-Fi signal can notice movement through a wall. That makes wireless sensing useful, but also tricky. Small changes in your body can bend the signal in ways that are hard to measure and even harder to classify. If you can turn those signal changes into reliable data, you can build a project that feels like a real engineering tool.

What Is It?

Wi-Fi channel state information, or CSI, describes how a wireless signal changes as it travels from one device to another. Walls, furniture, and people all bend that signal a little. Think of it like ripples in a pond. If you drop a pebble into the water, the ripples change when they hit a rock. CSI records those changes, so a model can guess whether someone is walking, standing still, breathing, or falling.

Your project would use a Wi-Fi device, such as an ESP32-S3, to collect those signal patterns. Then you would train a small machine learning model, such as a tiny transformer, to sort patterns into labels. A transformer is a model that looks for relationships across a sequence of data points, which helps when the signal changes over time. The privacy angle matters too. Unlike a camera, Wi-Fi sensing can detect presence and motion without storing images.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear signal pipeline from sensing to classification. You can change one part at a time, such as the activity, the distance, the wall material, or the number of people in the room, and measure how accuracy changes. The project connects to real problems in elder care, home monitoring, and privacy-friendly security. You also get room to learn RF sensing, data cleaning, feature design, and model evaluation, which makes the work feel real, not just like a demo.

Research Questions

• How does the activity type affect Wi-Fi CSI classification accuracy for walking, sitting, falling, and breathing patterns?
• What is the effect of wall material on through-wall sensing accuracy?
• Does the distance between transmitter and receiver change the model's ability to detect occupancy?
• To what extent does adding more training data improve fall detection compared with breathing-rate detection?
• Which preprocessing method gives the best classification results for noisy CSI traces?
• What is the effect of using a tiny transformer instead of a simpler model such as logistic regression or a small CNN?
• How does the number of people in the room change false positive rates for occupancy detection?

Basic Materials

• ESP32-S3 board with Wi-Fi support and USB cable.
• Second Wi-Fi-capable board or router for a transmitter-receiver setup.
• Laptop with Python installed.
• Tripod, tape, or desk stand for fixed device placement.
• Notebooks or spreadsheet for labeling each trial.
• Measuring tape for consistent spacing.
• Objects to mimic different wall materials, such as plywood, drywall, or a door.
• Stopwatch or phone timer for trial timing and labeling.

Advanced Materials

• ESP32-S3 boards or other CSI-capable Wi-Fi hardware.
• Access point and receiver setup with stable antennas.
• Laptop or workstation with Python, GPU access if available.
• External antennas with known gain for repeatable experiments.
• Spectrum analyzer or network analyzer for validation, if available.
• Environmental sensors for ground truth, such as a contact or pressure sensor for occupancy checks.
• Camera only as a separate reference tool, if your protocol and privacy rules allow it.
• Printed floor plan or room map for consistent geometry.

Software & Tools

• Python: Cleans CSI data, trains models, and plots accuracy across conditions.
• NumPy: Organizes signal arrays and supports fast data math.
• Pandas: Stores trial labels, metadata, and experiment logs.
• scikit-learn: Tests baseline classifiers and compares them with deeper models.
• TensorFlow Lite: Helps you build or test a tiny transformer for edge-friendly inference.
• ImageJ: Can help inspect plots or exported signal images when you convert CSI traces into visuals.

Experiment Steps

1. Define the sensing task you will study first, such as occupancy, breathing, or falls.
2. Choose one room setup and keep the transmitter, receiver, and subject positions fixed while you build your baseline.
3. Decide how you will label each trial so your training data matches the real activity, not just the signal shape.
4. Build a clean comparison plan that tests one factor at a time, such as wall type, distance, or number of people.
5. Select a simple baseline model before you try a tiny transformer, so you can measure whether the complex model really helps.
6. Plan your evaluation metrics, such as accuracy, precision, recall, and false alarm rate, before you collect the full dataset.

Common Pitfalls

• Changing device placement between sessions, which makes the CSI pattern shift for reasons unrelated to the activity.
• Collecting too few examples of rare events like falls, which leaves the model unable to learn them.
• Mixing labels during breathing trials, which happens when the subject moves or talks and the signal no longer matches the target.
• Testing only on the same room layout used for training, which hides how badly the system fails in a new space.
• Comparing a complex model to a weak baseline, which makes the transformer look better than it really is.

What Makes This Competitive

A stronger project would test more than one room, more than one wall type, and more than one person pattern. You could also compare a tiny transformer against simpler baselines and report where each model breaks down. Careful error analysis helps a lot, especially if you separate false alarms from missed detections. If you can show that your system keeps working under realistic changes, your project looks much closer to real engineering research.

Project Variations

• Use the same CSI setup to compare empty-room detection against person-count estimation.
• Swap fall detection for breathing-rate estimation and test how motion near the subject affects accuracy.
• Compare a tiny transformer with a CNN and a classical machine learning model on the same CSI dataset.

Learn More

• NIH PubMed: Search review articles on Wi-Fi sensing, radar sensing, and activity recognition to find background papers and related methods.
• IEEE Xplore: Search for papers on channel state information, through-wall sensing, and human activity recognition.
• MIT OpenCourseWare: Find free courses on signals, systems, probability, and machine learning to strengthen your analysis.
• NASA NTRS: Search for technical reports on wireless sensing, signal processing, and classification methods.
• arXiv: Search for preprints on Wi-Fi CSI sensing and tiny transformer models to see current research directions.
• GitHub: Search public repositories for ESP32 CSI collection scripts and open-source activity-recognition pipelines.