Smooth Cruize

We won!

Inspiration

Road damage is more than an inconvenience, it causes accidents, vehicle damage, and expensive maintenance delays. We were inspired by how many potholes go unreported or are reported too late, and wanted to build a simple way to turn everyday road footage into fast, useful signals for safer streets.

What it does

Smooth Cruize analyzes driving footage and automatically detects potholes using computer vision.

It then:

Highlights detections in processed video
Extracts short context clips around each pothole event
Organizes outputs so teams can quickly review hazards without manually scrubbing through full-length footage

How we built it

We built Smooth Cruize as a full-stack project:

Model inference: YOLO-based pothole detection pipeline in Python.
Video processing: OpenCV for frame-level processing and clip generation.
Batch workflow: A folder-based processor that scans multiple videos, detects potholes, merges nearby detections, and exports clean clips.
Backend: FastAPI service to expose processing logic.
Database and Authentication: Supabase
Frontend: Next.js app for a user-facing interface and workflow control.

Challenges we ran into

Detection quality vs speed: Balancing confidence thresholds, image size, and frame skipping to keep results accurate and efficient.
Noisy detections: Handling false positives and class-label inconsistencies across model outputs.
Clip quality: Generating useful pre/post-event context while avoiding duplicate or overlapping clips.
Video compatibility: Working across multiple input formats and inconsistent FPS metadata.
End-to-end integration: Connecting model, backend, and frontend into a smooth demo flow under hackathon time pressure.

Accomplishments that we're proud of

Built a working end-to-end system from raw video to actionable pothole clips.
Automated a tedious manual review process into a few simple steps.
Shipped a practical civic-tech concept with clear real-world utility.
Provide a clean workflow to monitor pothole hazards on the road

What we learned

Real-world video AI is heavily dependent on preprocessing and postprocessing, not just model choice.
Small tuning changes can dramatically impact precision and accuracy.
Clear output design (short, contextual clips) matters as much as detection itself.

What's next for Smooth Cruize

Improve model robustness with more diverse road-condition training data.
Add severity scoring (size/depth proxy) and prioritization of high-risk potholes.
Integrate public transport GPS/time metadata to map pothole locations automatically.
Support edge-friendly deployment for near real-time detection in the field.
Add more functionality to be able to detect a larger variety of road hazards.