A 24-hour hackathon build that reads sign language from your webcam

What it does The project captures live video through a webcam, isolates hand gestures with OpenCV, and classifies them using a Keras CNN trained on 44 characters of American Sign Language. It acts as a local translator that converts static ASL signs into text. The authors claim the model achieves above 95 percent accuracy on that limited vocabulary.

The interesting bit The entire pipeline—from calibrating a hand histogram to collecting training data, augmenting images by flipping them, and training the classifier—was built during a single 24-hour hackathon and won HackUNT-19. It is less a polished product than a complete proof-of-concept that happens to work end-to-end.

Key highlights

• Built in 24 hours at the University of North Texas HackUNT-19 and won first place.
• Recognizes 44 ASL characters using a custom CNN built with Keras and TensorFlow.
• Ships with a full data-collection toolkit for histogram calibration, gesture capture, image augmentation, and dataset splitting.
• Runs locally using only a webcam and a Python environment.
• The team published a 2025 preprint on Research Square and links to a follow-up advanced repository.

Caveats

• The vocabulary is limited to 44 ASL characters; the README explicitly lists expanding it as future work.
• The pipeline requires manual histogram calibration and gesture collection unless you rely entirely on the authors’ pre-built datasets.
• The project is marked as finished, and active development appears to have moved to a separate linked repository.

Verdict A useful reference if you want to see how a full gesture-recognition pipeline fits together with OpenCV and Keras. Look elsewhere if you need a maintained, cloud-deployed API or support for full ASL sentences rather than isolated characters.