Roadside Vehicle Classification With Microphones

ISEF Category: Embedded Systems

Subcategory: Signal Processing  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A single microphone can hear more than noise. It can pick out tires, engines, and sirens. That means one cheap sensor may classify vehicles in real time without a camera. You can turn that idea into a strong engineering project.

What Is It?

This project asks whether a computer can identify vehicles from sound alone. Think of it like recognizing friends by voice instead of face. A transformer is a model that learns patterns in a signal, and an MCU is a small microcontroller that can run code on its own.

Your core data comes from roadside audio. Cars, trucks, motorcycles, and emergency vehicles each leave different acoustic fingerprints. Tires, engines, road surface, and speed all shape the sound. Your job is to see whether a model can learn those patterns and keep working when the environment gets messy.

Hard-negative mining means you deliberately collect confusing examples. For this topic, that might mean ambulances, fire trucks, police vehicles, loud motorcycles, and buses that sound similar to your target classes. Those samples make the model work harder during training. They also help you test whether the system can handle the cases that matter most.

Why This Is a Good Topic

This makes a strong science fair project because you can measure real performance, not just build a demo. You can compare different models, test how well they run on small hardware, and see how extra confusing samples change accuracy. The work connects to traffic monitoring, smart intersections, road safety, and emergency response. You can learn signal processing, machine learning, embedded deployment, and experimental design in one project.

Research Questions

• How does adding hard-negative examples affect emergency-vehicle classification accuracy?
• What is the effect of using transformer features instead of hand-crafted audio features on vehicle-classification performance?
• Does roadside microphone placement change classification accuracy across vehicle types?
• To what extent does background traffic noise reduce model performance on emergency vehicles?
• Which vehicle classes are most often confused when the model runs on an MCU instead of a laptop?
• How does training with tire and engine signatures separately compare with training on full audio clips?

Basic Materials

• Single-board microcontroller with edge-ML support
• External USB microphone or microphone module
• Laptop for data labeling and model training
• Headphones for audio review
• Quiet test road or parking area with safe access
• Spreadsheet software for logging samples
• Free audio editor for clipping and labeling recordings
• Tripod or clamp for fixed microphone placement
• Measuring tape for consistent sensor distance
• Printed labels or notebook for metadata.

Advanced Materials

• Microcontroller development board with DSP or ML acceleration
• Calibrated reference microphone
• Audio interface or preamp
• Laptop with Python and training libraries
• External SSD for audio datasets
• Signal generator for testing input pipeline
• Access to a controlled roadway or test track recordings
• Accelerometer or GPS module for metadata alignment
• Acoustic calibration source
• Optional weather sensor for context variables.

Software & Tools

• Audacity: Trims recordings, checks noise levels, and helps you spot bad clips.
• Python: Cleans audio, extracts features, and trains comparison models.
• ImageJ: Not used here.
• scikit-learn: Builds baseline classifiers so you can compare them with your transformer.
• PyTorch: Trains and tests the transformer model on audio data.
• Edge Impulse: Helps you prototype edge models and check whether they fit on small hardware.

Experiment Steps

1. Define the exact classes your system will recognize, and decide which vehicles count as hard negatives.
2. Plan a recording setup that keeps microphone position, sample labeling, and site conditions as consistent as possible.
3. Build a baseline model first, so you can measure whether the transformer actually improves performance.
4. Design a comparison that separates normal training from hard-negative mining, then decide which metrics will prove the difference.
5. Check whether the model still works after deployment on the MCU, and compare its output to the laptop version.
6. Plan a failure analysis that groups the clips the model confuses most, so you can explain why those errors happen.

Common Pitfalls

• Mixing different microphone positions across sessions, which makes the model learn location instead of vehicle sound.
• Using too few emergency-vehicle clips, which can make the classifier fail on the rare class you care about most.
• Labeling trucks, buses, and vans loosely, which blurs class boundaries and weakens the confusion analysis.
• Training only on clean daytime recordings, which makes the model collapse when traffic noise, wind, or sirens appear.
• Skipping MCU deployment testing, which leaves you with a model that looks good on a laptop but fails on embedded hardware.

What Makes This Competitive

A strong version of this project does more than report accuracy. It compares model types, tests deployment limits, and explains where the system fails. You can make it stand out by using a careful confusion analysis, by showing how hard-negative mining changes emergency-vehicle recall, or by measuring accuracy under different noise conditions. If you can connect those results to a real roadside use case, the project looks much stronger.

Project Variations

• Test whether the system works better with one microphone or a small microphone array for the same vehicle classes.
• Compare daytime, nighttime, and wet-road recordings to see how environment changes acoustic classification.
• Swap emergency vehicles for construction vehicles or buses to study which confusing class pairs hardest-negative mining helps most.

Learn More

• PubMed: Search review articles on acoustic classification, environmental sound recognition, and machine listening methods.
• IEEE Xplore: Search for papers on edge deployment of audio classifiers and transformer models for embedded systems.
• MIT OpenCourseWare: Look for free courses on signal processing, machine learning, and embedded systems design.
• NOAA: Use weather and wind resources to think about how outdoor conditions affect roadside audio recordings.
• NASA: Search for free resources on signal analysis and classification methods that connect to remote sensing.
• USGS: Review field data collection guidance and sensor placement ideas that translate well to outdoor experiments.