Ultra-Low-Power Gesture Wakeword for Smart Glasses

ISEF Category: Embedded Systems

Subcategory: Microcontrollers  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Your smart glasses cannot stay awake all day if they burn power every second. That is why a tiny gesture wakeword matters. The trick is to wake the device only when your hand motion says, “pay attention.”

What Is It?

This project studies a tiny always-on gesture detector that runs on a microcontroller, not a phone or laptop. Think of it like a quiet guard at a door. The guard listens for a simple motion pattern, then wakes the larger system only when needed.

A Cortex-M0+ is a small, low-power chip found in many embedded devices. A depthwise CNN, or convolutional neural network, is a pattern-matching model that can learn motion shapes from accelerometer data. Int4 quantization means the model stores numbers in a very compact format, which cuts memory use and often lowers power draw. Your job is to see how far you can shrink the model while still spotting gestures well.

Why This Is a Good Topic

This makes a strong science fair topic because you can measure both engineering tradeoffs and user impact. You can test accuracy, latency, memory use, and battery drain, then ask how each one changes when you shrink the model or change the wake logic. That connects directly to wearables, assistive tech, and always-on sensing. A student can learn signal processing, embedded AI, and experimental design from one project.

Research Questions

• How does int4 quantization affect gesture classification accuracy compared with higher-precision models?
• What is the effect of model size on average current draw during always-on listening?
• Does adding a wakeword stage reduce total energy use compared with continuous gesture classification?
• To what extent does sensor placement on the glasses frame change gesture recognition accuracy?
• Which gesture classes are most often confused when the model runs on a Cortex-M0+?
• How does sampling rate affect the tradeoff between accuracy and power consumption?

Basic Materials

• Cortex-M0+ development board with accelerometer support
• USB cable and computer
• Breadboard and jumper wires
• Accelerometer breakout board with I2C or SPI interface
• Multimeter with microamp or milliamp range
• Logic analyzer or oscilloscope, if available
• Rechargeable battery pack or bench power supply
• Notebook for test logs
• Tape or clips for repeatable sensor placement

Advanced Materials

• Cortex-M0+ microcontroller board with low-power mode support
• High-resolution power analyzer or source measure unit
• External 3-axis accelerometer with stable mounting hardware
• Smart-glasses frame or head-mounted test rig
• JTAG or SWD debugger
• Temperature probe, if thermal drift matters
• Reference microcontroller board for baseline comparisons
• Small vibration source for repeatable motion tests
• Calibration weights or mounting jig for sensor alignment

Software & Tools

• Python: Cleans accelerometer data, runs statistical tests, and plots accuracy versus power.
• TensorFlow Lite Micro: Deploys a tiny CNN model to a microcontroller for edge inference.
• Edge Impulse: Helps collect motion data, train lightweight models, and compare embedded performance.
• Jupyter Notebook: Keeps your analysis, charts, and parameter sweeps organized in one place.
• ImageJ: Can help inspect timing plots or screenshots when you need frame-by-frame verification.

Experiment Steps

1. Define the exact gesture set you want the device to recognize and decide which gestures count as wake events.
2. Choose the performance metrics you will track, including accuracy, false wake rate, latency, memory use, and current draw.
3. Build a baseline model first, then plan one change at a time, such as quantization level, model depth, or sampling rate.
4. Design controls that separate true motion recognition from noise, drift, and ordinary head movement.
5. Plan a fair test protocol for repeated trials across users, mounting positions, and battery conditions.
6. Decide how you will compare power savings from a wakeword stage against a continuous always-on model.

Common Pitfalls

• Training on one person's motion data, which makes the model fail on new users.
• Measuring current with a meter setup that cannot capture short bursts, which hides real power spikes.
• Changing sensor placement between trials, which makes the accelerometer data look inconsistent.
• Reporting accuracy without false wake rate, which ignores the cost of triggering by mistake.
• Testing only clean, scripted gestures, which overstates performance in real use.

What Makes This Competitive

A stronger version of this project goes beyond a simple accuracy test. You can compare several model sizes, quantify the power cost of each design choice, and show the point where smaller models stop helping. You can also test across users, placements, and motion styles, then use confusion matrices and significance tests to show what really changes. That kind of analysis looks much closer to real embedded product research.

Project Variations

• Test the same wakeword idea on a wristband instead of smart glasses to compare how body position changes motion signatures.
• Swap the accelerometer-only model for a fused accelerometer-plus-gyroscope model and compare power against accuracy.
• Compare a continuous classifier with a two-stage wakeword plus gesture model to measure total battery savings.

Learn More

• MIT OpenCourseWare: Search for embedded systems, machine learning for small devices, and microcontroller courses with lecture notes and assignments.
• TensorFlow Lite for Microcontrollers: Read the official documentation for tiny ML deployment on resource-limited hardware.
• Edge Impulse Docs: Find free guides on motion classification, model training, and embedded benchmarking.
• NIH PubMed: Search review articles on wearable sensing, activity recognition, and low-power edge AI.
• IEEE Xplore: Search recent journal articles on ultra-low-power gesture recognition and quantized neural networks.
• NASA Technical Reports Server: Search for signal processing and low-power sensing reports that often include practical embedded methods.