Phone-Based Distracted Driving Detection App

ISEF Category: Systems Software

Subcategory: Mobile Apps  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A split-second glance at a phone can raise crash risk fast. Your phone already has a camera, motion sensors, and a brain for pattern matching. That makes it a strong candidate for spotting risky driving habits before they turn into a crash. You can test whether a battery-saving app can catch those moments without draining the phone.

What Is It?

This project asks a simple question with a hard answer: can a phone detect small driving mistakes from its own sensors? The front camera can watch for face angle or head pose, and the IMU, which measures motion and rotation, can catch sudden swerves, jerks, and lane drift. Think of it like giving the phone two senses, sight and balance, then asking whether those senses together can spot danger.

The real challenge is not only accuracy. Your app also has to save battery, because a driving safety app cannot die halfway through a trip. That means you can test a duty-cycled pipeline, which turns sensing or inference on and off in planned bursts instead of running all the time. You can compare full-time detection, camera-only detection, IMU-only detection, and mixed modes to see which design gives the best tradeoff.

Why This Is a Good Topic

This is a strong science fair topic because you can measure real performance, not just build a demo. You can test accuracy, false alarms, latency, and battery drain, all with clear numbers. The project connects to road safety, human attention, and mobile system design. A student can also learn data cleaning, model evaluation, sensor fusion, and power-aware app design without needing a university lab.

Research Questions

• How does using camera data alone compare with IMU data alone for detecting distracted-driving micro-behaviors?
• How does fusing front-camera features with IMU features change detection accuracy compared with either sensor by itself?
• What is the effect of duty-cycled inference on battery use while keeping detection performance acceptable?
• To what extent does the model miss brief phone-glance events when inference runs less often?
• Which feature set gives the best balance of false positives, false negatives, and battery drain?
• How does changing the decision threshold affect recall for sudden lane drift events?

Basic Materials

• Smartphone with front camera and IMU sensors
• Laptop or desktop computer for model training and evaluation
• Public distracted-driving dataset, such as the StateFarm distracted-driving dataset
• Python installed on a computer
• Jupyter Notebook or Google Colab
• Spreadsheet software for tracking experiments
• USB cable for file transfer and device charging
• External battery pack for battery-drain tests

Advanced Materials

• Smartphone with developer access for on-device testing
• Laptop or workstation with a GPU for faster model training
• Public distracted-driving dataset, such as the StateFarm distracted-driving dataset
• Android Studio or Xcode for app prototyping
• Python environment with machine-learning libraries
• OpenCV for video feature extraction
• TensorFlow Lite or PyTorch Mobile for deployment testing
• Power monitor or battery logging tools for energy measurements

Software & Tools

• Google Colab: Gives you free cloud compute for training and model comparison when your laptop is slow.

Experiment Steps

1. Define the exact risky behaviors you want to detect, and decide whether you will classify single frames, short clips, or sensor windows.
2. Choose a baseline model first, then plan a stronger version that adds either sensor fusion or duty-cycled inference.
3. Map the dataset labels to your target classes, and decide how you will split training, validation, and test data so the same driver or trip does not leak across splits.
4. Design controls that separate visual information from motion information, so you can test camera-only, IMU-only, and combined pipelines fairly.
5. Plan evaluation metrics that matter for safety and phone use, including recall, precision, latency, and battery cost.
6. Decide how you will summarize tradeoffs, so you can argue which design makes the best mobile safety app, not just the highest score.

Common Pitfalls

• Training and testing on overlapping drivers or trips, which makes the model look better than it really is.
• Ignoring class imbalance, which can hide poor detection of rare dangerous behaviors.
• Using camera frames with inconsistent cropping or lighting, which makes visual features noisy.
• Measuring battery use only on a laptop simulation, which misses the cost of running the app on a real phone.
• Reporting only accuracy, which can hide false alarms and missed dangerous events.

What Makes This Competitive

A competitive project would go beyond a basic classifier. You could compare multiple sensor-fusion designs, then test how well each one works under real mobile limits like battery drain and delayed inference. Strong entries also use clean splits, solid error analysis, and safety-focused metrics such as recall on rare events. If you add a careful comparison between performance and power use, your project starts to look like real systems research.

Project Variations

• Focus on detecting phone glances only, then compare front-camera features against head-pose estimates from the same video.
• Focus on lane-drift detection from IMU signals alone, then test whether shorter sensing windows still catch the event.
• Focus on energy-aware mobile design by comparing always-on inference with duty-cycled inference across the same model.

Learn More

• StateFarm Distracted Driver Detection Dataset: Search for the public Kaggle dataset and read the label definitions, sample structure, and evaluation notes.
• PubMed: Search review articles on distracted driving, attention, and crash risk to ground your problem in real safety research.
• NIH NCBI Bookshelf: Read free background chapters on sensors, machine learning basics, and mobile health system design.
• MIT OpenCourseWare: Search for free courses on machine learning, computer vision, or mobile systems to learn the core methods behind your app.
• TensorFlow Lite Guide: Read the free official documentation to learn how lightweight on-device inference works.
• NOAA National Highway Traffic Safety resources: Search federal traffic safety data and reports for background on distracted-driving harms.