Smart Cane Obstacle Alerts for Blind Navigation

ISEF Category: Biomedical Engineering

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

A curb can look flat to you and become a real hazard to someone using a cane. That gap is why smart mobility tools matter. Your project can turn radar and vision data into fast, useful warnings. You get to test whether a cane can help people notice danger before they step into it.

What Is It?

This project explores a smart white-cane system that spots ground-level hazards and speaks them to the user through bone-conduction audio. Bone-conduction earphones send sound through the bones near your ear instead of blocking your ears, so the user can still hear traffic and people around them.

The core idea is sensor fusion, which means combining data from more than one sensor so the device can make a better guess. In this case, a camera-based object detector like YOLO can recognize shapes, while mmWave radar can help detect distance and motion even when lighting is bad. Think of it like two friends describing the same scene. One sees the shape, and the other confirms how far away it is.

Why This Is a Good Topic

This is a strong science fair topic because you can test clear, measurable outcomes, like detection accuracy, false alarms, reaction time, and navigation safety. It connects to a real access problem for people who are blind or have low vision, so your work has a direct human impact. You can also scale the project to match your resources, from a simple obstacle-classification prototype to a full navigation trial with statistical analysis.

Research Questions

• How does sensor fusion affect the accuracy of curb, pothole, and stair detection?
• What is the effect of outdoor lighting on camera-only, radar-only, and fused obstacle detection?
• Does bone-conduction audio help users notice alerts faster than a phone speaker or buzzer?
• To what extent does obstacle distance change the model's ability to classify ground-level hazards?
• Which obstacle type, curb, pothole, or stairs, produces the highest false-negative rate?
• How does changing the sensor mounting angle affect detection accuracy during walking trials?

Basic Materials

• White cane or cane-length handle for mounting components.
• LD6001 mmWave radar module or similar short-range radar sensor.
• Small camera module compatible with your processor.
• Single-board computer or microcontroller with enough processing power for on-device inference.
• Bone-conduction earphone or headset.
• Portable battery pack.
• Mounting brackets, zip ties, and foam padding.
• Measuring tape for trial distances.
• Sidewalk-safe obstacle props, such as foam curb blocks, shallow boxes, and marked stair edges.
• Consent forms and a data sheet for trial logging.

Advanced Materials

• LD6001 mmWave radar module with documented interface.
• Camera module with adjustable field of view.
• NVIDIA Jetson or similar edge AI computer.
• Bone-conduction audio output device.
• IMU for cane motion tracking.
• GPS logger for route mapping.
• 3D-printed mounting system for sensor alignment.
• Reflective or textured obstacle props for controlled testing.
• Laptop for model training, annotation, and evaluation.
• Calibration targets for distance and angle checks.

Software & Tools

• Ultralytics YOLO: Trains and tests object detection models for curb, pothole, and stair recognition.
• ImageJ: Measures image quality, obstacle size, and annotation consistency in sample frames.
• Python: Handles sensor fusion, trial logging, and statistical analysis.
• OpenCV: Processes camera frames before they enter the detection model.
• Google Colab: Gives you a free cloud notebook for model experiments when your laptop is slow.

Experiment Steps

1. Define the hazard classes you will detect and decide how each one will be labeled in your data.
2. Choose the one performance metric that matters most, such as classification accuracy, missed hazards, or alert latency.
3. Plan a baseline system first, then compare it with a fused sensor version so you can measure the added value of radar.
4. Design your data collection route, obstacle spacing, and trial conditions so each test stays repeatable.
5. Build controls that separate sensor error from user error, lighting changes, and mounting issues.
6. Set up your analysis plan before testing so you can compare confusion matrices, precision, recall, and navigation outcomes.

Common Pitfalls

• Training on too few examples of curbs, potholes, and stairs, which makes the model memorize the test set instead of learning hazard features.
• Mixing obstacle sizes and shapes without clear labels, which turns the class boundaries into noise.
• Testing only in one lighting condition, which hides how much the camera fails at dusk, glare, or shadow.
• Mounting the radar and camera at slightly different angles each time, which breaks sensor alignment during fusion.
• Reporting only model accuracy and ignoring false negatives, even though missed hazards matter most for safety.

What Makes This Competitive

A competitive version of this project would do more than build a working prototype. You would compare sensor combinations, test them across different environments, and report the tradeoffs with clear statistics. Strong projects also explain failure cases, not just success cases. If you can show how one design choice improves safety or lowers missed detections, your work becomes much stronger.

Project Variations

• Test whether the cane works better on curbs, potholes, or stairs by changing only the obstacle class.
• Compare camera-only detection against radar-only detection, then measure how much fusion improves missed-hazard rates.
• Study whether vibration feedback, speech output, or bone-conduction audio leads to faster user response during navigation trials.

Learn More

• NIH PubMed: Search review articles on assistive technology, blind navigation, and sensor fusion for mobility aids.
• IEEE Xplore: Search for peer-reviewed papers on smart canes, radar sensing, and edge AI for assistive devices.
• NASA Open Source Tools and Open Data: Browse free image-processing and data-analysis resources that can support computer vision work.
• MIT OpenCourseWare: Find lectures on machine learning, signal processing, and embedded systems that help with model design.
• NOAA and USGS data portals: Use environmental and terrain data sources to think about outdoor testing conditions and ground surface variation.