RC Car Vision Lane Keeping Under Hard Conditions

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Ground Vehicle Systems  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

A car can lose its lane with one bright reflection. Your model can too. That makes this project more than a toy, because real roads have glare, shadows, and messy markings. If your RC car can stay centered when the floor gets ugly, you have a real engineering problem to solve.

What Is It?

This project asks a simple question with a tricky answer, can a small camera and a tiny neural network keep an RC car inside a lane without human help? The car looks at the floor, finds the lane markings, and decides how to steer. Think of it like teaching the car to follow a ribbon on the ground, except the ribbon can fade, break, or disappear under light changes.

The vision system usually has three parts. The camera captures the scene, the model finds lane features, and the controller turns that visual signal into steering commands. A tiny CNN, or convolutional neural network, is a compact pattern-finding model that can learn shapes, edges, and lane positions from images. On an ESP32-S3, that model has to work with limited memory and speed, so every design choice matters.

Your research twist is robustness. You are not just asking whether the car can follow a clean lane. You are asking how much performance drops when glare, shadows, and partial tape removal change the visual scene. That turns a demo into a real systems project, because you are testing perception, control, and environmental stress together.

Why This Is a Good Topic

This is a strong science fair topic because you can test it in a repeatable way, and you can measure clear numbers like lane centering error, track completion rate, and recovery time after a bad frame. It connects to real problems in driver assistance, warehouse robots, and autonomous ground vehicles. You can also change one factor at a time, like lighting or occlusion, and see how the system responds. That gives you enough depth for a serious project without needing a full-size vehicle or a university lab.

Research Questions

• How does glare on the lane surface affect lane-centering error for a vision-only RC car??
• How does shadow width across the lane affect the car's ability to stay in its lane??
• How does partial occlusion of lane tape change track completion rate??
• What is the effect of different lane tape colors on model accuracy under the same lighting??
• To what extent does training the CNN with augmented glare and shadow images improve robustness??
• Which camera mounting angle gives the lowest steering error on a taped gym-floor track??

Basic Materials

• 1:10 RC car chassis with steering control.
• ESP32-S3 microcontroller board with camera support.
• OV5640 camera module or compatible small camera.
• USB cable and computer for flashing firmware.
• Tape for lane markings in at least two colors.
• Measuring tape or meter stick.
• Stopwatch or phone timer.
• Tripod or fixed mount for repeatable filming.
• Gym floor, hallway floor, or smooth indoor test surface.
• Bright desk lamp or portable LED light.
• Masking tape for marking test zones.
• Notebook or spreadsheet for data logging.

Advanced Materials

• 1:10 RC car chassis with steering and throttle control.
• ESP32-S3 development board with camera interface.
• OV5640 camera module with stable mounting hardware.
• External battery pack or regulated power supply for repeatable tests.
• Calibrated light meter or lux meter.
• Portable LED panels with adjustable angle.
• Neutral density film or diffusers for glare control.
• Printed calibration targets for camera alignment.
• High-contrast lane tapes in several widths and colors.
• High-speed phone camera or action camera for validation footage.
• Laptop for model training, analysis, and visualization.
• Optional inertial sensor or wheel encoder for motion comparison.

Software & Tools

• Arduino IDE: Uploads firmware to the ESP32-S3 and helps you test camera and steering code.
• Edge Impulse: Lets you train and export a tiny model that can run on small hardware.
• Python: Helps you clean data, analyze accuracy, and graph performance across conditions.
• ImageJ: Measures lane brightness, shadow coverage, and image contrast in sample frames.
• OpenCV: Helps you inspect frames, mark lanes, and compare how the car sees each test condition.

Experiment Steps

1. Define the performance metric you will use, such as centered lane travel, steering error, or completed laps without intervention.
2. Choose one lane-following approach to test first, then keep the hardware and controller fixed while you vary the environment.
3. Plan a controlled track layout with repeatable tape geometry, fixed camera height, and marked test zones for glare, shadows, and occlusion.
4. Build a training and validation image set that includes both clean lanes and stressed lanes, then separate your test data by condition.
5. Decide how you will compare baseline performance with robustness-focused training, such as augmented images or different lane widths.
6. Set up a scoring system that captures both success rate and failure mode, so you can explain why the car lost the lane.

Common Pitfalls

• Training on clean, centered lanes only, which makes the model fail as soon as the floor lighting changes.
• Letting the camera angle shift between runs, which changes the lane shape in every frame and breaks comparison.
• Testing glare, shadow, and occlusion all at once, which makes you unable to tell which stress caused the error.
• Using lane tape colors that blend into the gym floor, which lowers contrast and hides whether the model or the scene caused the failure.
• Measuring only whether the car finishes the track, which misses small steering errors that explain later crashes.

What Makes This Competitive

A stronger version of this project goes past simple success or failure. You can compare several stress conditions, quantify the drop in performance, and explain which visual features break first. You can also test whether data augmentation, different camera angles, or lane geometries improve generalization. That kind of careful analysis makes the project feel like systems engineering, not just a working robot.

Project Variations

• Test how the system handles different lane widths, then compare whether narrow lanes or wide lanes cause more steering drift.
• Swap the gym floor for a glossy hallway or matte track surface, then measure how floor reflectivity changes camera perception.
• Compare a tiny CNN with a simple classical vision pipeline, then see which method fails first under shadows and tape gaps.

Learn More

• MIT OpenCourseWare: Search for introductory computer vision and embedded systems courses that explain image pipelines and small-device deployment.
• PubMed: Search for review articles on autonomous driving perception, lane detection, and robustness under adverse visual conditions.
• NASA Image Analysis resources: Find free material on image processing, feature detection, and measurement from camera data.
• NIH 3D Print Exchange or related NIH visualization resources: Use free guidance on image-based measurement workflows and data handling.
• OpenCV Documentation: Read the official guides for image preprocessing, feature extraction, and camera frame handling.
• IEEE Xplore abstracts: Search for recent lane detection and small-model robotics papers, then read abstracts and open-access versions where available.