Phone Pothole Detection for Road Quality Mapping

ISEF Category: Embedded Systems

Subcategory: Signal Processing  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A phone can hear a pothole before you see it. Tiny shakes in the accelerometer and gyroscope can act like a fingerprint for bad pavement. If you can tell those patterns apart, you can help map rough roads block by block.

What Is It?

This project asks your phone to act like a road-quality sensor. Your phone has an IMU, short for inertial measurement unit. That means it can measure motion, tilt, and sudden jolts. When a bike, car, or cart hits a bump, the motion signal changes fast. Your job is to turn those wiggles into a clear yes-or-no or roughness score.

Wavelet decomposition breaks a signal into pieces at different time scales. Think of it like splitting a song into bass, drums, and vocals, except you are splitting motion into slow trends and sharp hits. A tiny transformer is a small machine learning model that looks for patterns across those pieces. You can compare that model against simpler baselines, like accelerometer-only rules or basic statistics, to see which one spots road anomalies better.

The final output can become a road-quality tag, similar to a map note that says a street feels rough. OpenStreetMap can store that kind of field data. So your project connects signal processing, mobile sensing, and civic mapping.

Why This Is a Good Topic

This is a good science fair topic because you can test it with real motion data and clear labels. You do not need a university lab, just a phone, a few routes, and a careful plan for collecting and comparing signals. The real-world link is strong, since bad roads affect safety, biking, accessibility, and vehicle wear. You can learn how features, model choice, and evaluation methods change accuracy.

Research Questions

• How does wavelet decomposition affect pothole detection accuracy compared with raw IMU signals?
• What is the effect of using accelerometer-only data versus accelerometer plus gyroscope data on road-anomaly classification?
• Does a tiny transformer outperform a rule-based baseline for detecting rough-road events from phone motion data?
• To what extent does device placement, such as pocket, backpack, or mounted holder, change detection performance?
• Which wavelet features best separate potholes, speed bumps, and normal pavement?
• How does class imbalance affect false alarms when you tag road quality from phone IMU data?
• To what extent do OpenStreetMap road-quality tags agree with your model’s predicted anomaly locations?

Basic Materials

• Smartphone with accelerometer and gyroscope sensors.
• Stable phone mount, bike mount, or car mount.
• Vehicle, bike, scooter, or walking route with known rough and smooth sections.
• Notebook or spreadsheet for route labels and ground truth notes.
• GPS tracking app with export option.
• Digital kitchen scale or ruler for basic setup checks, if needed.
• Computer for data cleaning, plotting, and model training.
• Charging cable and external battery pack for long collection sessions.

Advanced Materials

• Smartphone with access to raw IMU sampling at a known rate.
• External IMU or motion logger for comparison data.
• GPS receiver with higher location accuracy, if available.
• Edge device or microcontroller for on-device inference tests.
• Annotated road surface reference set for label checking.
• Cloud storage or local server for data uploads and backups.
• Reference wheel odometer or survey tool for route distance checks.
• Access to OpenStreetMap editing tools for quality tag experiments.

Software & Tools

• Python: Cleans sensor data, computes features, and trains baseline and transformer models.
• Pandas: Organizes time-stamped IMU, GPS, and label data.
• NumPy: Handles signal arrays and numerical feature calculations.
• SciPy: Supports signal filtering, peak finding, and wavelet-related preprocessing.
• ImageJ: Lets you inspect plotted signal images or spectrogram-style exports when you need a visual check.

Experiment Steps

1. Define one road-anomaly target, such as potholes, speed bumps, or rough pavement, so your labels stay consistent.
2. Choose a data collection setup that keeps the phone position fixed and records the same sensor channels every time.
3. Plan a labeling method that ties each motion burst to a known road segment or a mapped location.
4. Decide which signal versions you will compare, such as raw IMU, wavelet features, and simplified baseline features.
5. Build a fair evaluation plan with held-out routes, confusion matrices, and location-based error checks.
6. Design a mapping output format so your predictions can become road-quality tags that match OpenStreetMap style data.

Common Pitfalls

• Changing the phone mount between runs, which makes the vibration pattern look like a road effect when it is really a setup effect.
• Using GPS points alone to label potholes, which can miss the exact bump location because phone location drifts.
• Collecting too few rough-road examples, which makes the model overpredict the smooth class.
• Mixing walking, biking, and car data in one model without separating speed effects, which hides the road signal.
• Comparing models with different route sets, which makes accuracy numbers unfair and easy to misread.

What Makes This Competitive

A stronger version of this project will do more than spot bumps. You can compare multiple sensor inputs, test route transfer across different streets, and measure whether the model still works when speed or mount position changes. Good entries also use careful ground truth, not just a loose guess about where the pothole was. If you add location-aware analysis and a fair baseline, your project starts to look like real field sensing research.

Project Variations

• Test whether road-anomaly detection changes when the phone is mounted on a bike, held in a backpack, or fixed in a car.
• Compare wavelet features with spectrogram features to see which signal representation gives better classification.
• Focus on one city block and study how your predicted roughness tags match OpenStreetMap road-quality edits over time.

Learn More

• OpenStreetMap Wiki: Search the tags and mapping notes for road surface, smoothness, and quality-related fields.
• NASA Earthdata: Find open geospatial data and mapping workflows that help with location-based analysis.
• NOAA National Centers for Environmental Information: Use environmental data records and mapping references for location-aware projects.
• MIT OpenCourseWare, Signal Processing: Search lecture notes on time-frequency analysis, filtering, and feature extraction.
• PubMed: Search review articles on mobile sensing, pothole detection, and road condition monitoring.