Inspiration
A safety inspector that looks twice. Same-level falls cost $9.92 billion a year in U.S. direct workers' comp. In warehouses, stores, and kitchens, a spill or a blocked exit is waiting for someone to get hurt. A plain detector either guesses when evidence is borderline, or floods staff with alarms.
What it does
VisionCrew Inspector: OpenCV 5 picks up blur, exposure, spill colour, and exit edges. Then an agent decides from confidence bands, in bounded steps. Then it acts: verdict, marked-up image, and full decision trace. When unsure, it does not guess. It looks again, or asks a person. Version 2 is a 10-step fail-safe pipeline. A person is always the last step.
How we built it
Coded in Python, JavaScript, HTML, and CSS, using OpenCV 5, FastAPI, AWS Lambda, CloudWatch, and Docker. Try the interactive dashboard at the demo link on this page.
Results
329 out of 329 tests pass. 66 synthetic spills: version 1 let 26 slip through, version 2 let zero through. 1,008 stress-test images: recall went from 62.6% to 99.9%, that is 764 out of 765. Precision is 77.6%, more escalations, stated plainly. Real photos: 40 public-domain images, plus 8 blur and dark variants. Every spill, blocked exit, and unusable photo is caught. Zero hazards get auto-passed.
Limitations, stated plainly
These are classical detectors, laid out for one calibrated camera. No real-site accuracy is claimed. Live AWS deployment is still pending.
What is next
Next up: a deep-learning model, live video, and on-site calibration.
Author
By Gabriel Aicon.
Built With
- amazon-web-services
- cloudwatch
- css
- docker
- fastapi
- html
- javascript
- lambda
- opencv
- python
Log in or sign up for Devpost to join the conversation.