Phone IMU Walking Assist App for Navigation

ISEF Category: Systems Software

Subcategory: Mobile Apps  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Your phone already feels tiny, but its motion sensors can act like a body-level radar. If you can read those signals well enough, you can tell when a walker hits a curb, starts stairs, or reaches a bus stop. That opens a privacy-friendly way to help blind and low-vision users navigate without a camera. The hard part is teaching a phone to notice the difference between normal walking and a real landmark.

What Is It?

This project tests whether a phone’s inertial measurement unit, or IMU, can recognize walking patterns linked to nearby landmarks. An IMU measures acceleration and rotation. Think of it like a tiny balance sense inside the phone. When your steps change because the ground changes, the phone records those changes as patterns in motion data.

TinyML means machine learning made small enough to run on a phone. Instead of sending data to a server, the app runs the model on the device. That helps with privacy and can save battery. Your goal is to see whether the model can separate normal walking from events like curb approach, stair climbing, or stopping at a bus stop by using only sensor data.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear question, collect your own data, and measure accuracy with real numbers. It connects to accessibility, privacy, and mobile computing, which gives the project real-world value. You can also compare models, sensor windows, and feature sets, so you have many ways to make the work your own. A student can learn signal processing, basic machine learning, and field validation without needing a camera or a special robot.

Research Questions

• How does using only phone IMU data affect the accuracy of detecting curbs, stairs, and bus stops?
• What is the effect of sensor window length on TinyML classification accuracy for walking landmarks?
• Does adding gyroscope data improve detection more than using accelerometer data alone?
• To what extent does walking speed change the model's ability to identify curb and stair events?
• Which model type, such as decision tree, random forest, or small neural network, gives the best tradeoff between accuracy and phone-friendly size?
• What is the effect of training on data from one route and testing on a new route on classification performance?
• To what extent do phone placement and carrying style change the gait signatures the model learns?

Basic Materials

• Smartphone with accelerometer and gyroscope sensors
• Second smartphone or stopwatch for time stamps
• Notebook or spreadsheet for labeling routes and events
• Outdoor walking route with safe curb, stair, and stop locations
• Measuring tape or route map for documenting landmarks
• Optional wearable phone pouch or armband for consistent placement
• Consent forms and safety plan for human participants
• Digital kitchen scale for checking any carried load conditions.

Advanced Materials

• Smartphone or development board with raw IMU logging access
• Laptop with Python and Jupyter notebooks
• TensorFlow Lite or Edge Impulse for TinyML prototyping
• Bluetooth IMU logger, if the phone app needs a comparison sensor
• Video timestamp reference system for ground-truth labeling
• Annotated route map with landmark coordinates
• Statistical analysis software for model comparison
• Access to a university accessibility or human factors lab for pilot testing.

Software & Tools

• Python: Cleans sensor data, builds features, and compares model performance.
• Jupyter Notebook: Helps you explore time-series plots and document each analysis step.
• TensorFlow Lite: Lets you test small models that can run on a phone.
• Edge Impulse: Supports sensor-based TinyML workflows and quick model prototyping.
• ImageJ: Not for image analysis here, but useful only if you later add visual route validation data.

Experiment Steps

1. Define the exact walking events you want the app to detect, and write clear labels for each one.
2. Plan a data collection route that gives you repeated examples of each event under controlled conditions.
3. Choose one sensor setup first, then decide how you will segment raw IMU data into model inputs.
4. Build a baseline classifier before you try a smaller TinyML version, so you know whether the problem is learnable.
5. Design controls for phone placement, walking speed, and route differences, so you can test generalization instead of memorization.
6. Set your evaluation plan before model training, including how you will measure accuracy, false alarms, and missed events.

Common Pitfalls

• Mixing up curb, stair, and stop labels during data collection, which poisons the training set.
• Letting phone placement vary between pockets, hands, and bags, which changes the IMU signal more than the landmark does.
• Training and testing on the same route, which inflates accuracy and hides poor real-world performance.
• Using a model that is too large for on-device inference, which defeats the point of a phone-only app.
• Ignoring ground-truth timing errors, which makes the model look wrong even when the sensor pattern is useful.

What Makes This Competitive

A strong version of this project does more than build a classifier. It asks whether the model still works across different routes, users, and phone placements. It also compares accuracy against battery use, model size, and false alarm rate, not just one score. If you add careful field validation and a clean analysis of failure cases, your project starts to look like real applied research.

Project Variations

• Test whether the same IMU model works for detecting bike crossings, crosswalk stops, or elevator entry instead of stairs and curbs.
• Compare phone pocket placement, hand-held placement, and backpack placement to see which gives the most stable gait signatures.
• Build one version that uses raw IMU windows and another that uses handcrafted features, then compare accuracy and phone efficiency.

Learn More

• NIH PubMed: Search for review articles on wearable inertial sensors, gait analysis, and mobile accessibility.
• NASA Open Source Software Catalog: Explore examples of sensor processing, edge computing, and on-device inference projects.
• MIT OpenCourseWare: Search for free courses on machine learning, signal processing, and embedded systems.
• IEEE Xplore: Read abstracts and papers on TinyML, inertial sensing, and mobile assistive technology through your school or public library access.
• TensorFlow Lite documentation: Learn how small neural networks run on phones and other edge devices.
• USGS or NOAA mapping tools: Use public maps and terrain data to plan and document outdoor test routes.