A camera supplies input to the real-time sign-detection pipeline.
What I built
Camera input runs through a custom YOLO model on Raspberry Pi 5 for real-time sign detection.
I trained the model with PyTorch on Google Colab GPUs using open-source ASL datasets, then optimized deployment for the target device.
Purpose & audience
Make American Sign Language detection more accessible in educational settings.
Built with students with hearing impairments and educators using affordable classroom hardware in mind.
Under the hood.
Follow the architecture, one connection at a time.
SignLect / architecture
A custom YOLO model trained with PyTorch on open-source ASL datasets detects signs in camera input.
The model is optimized for Raspberry Pi 5, bringing the detection pipeline onto affordable edge hardware.
The device returns detected signs. This project focuses on sign recognition; full language translation is a broader problem.
Scroll to follow the connections. Select a step to take a closer look.
Decisions & tradeoffs
Model compression and hardware-aware deployment shaped the project, with responsiveness and affordability central to the classroom use case.
Training on GPU hardware and running on a Raspberry Pi impose different constraints. Inference needs to be evaluated on the device where the system will actually run.
Outcomes & limitations
Built and optimized an edge-deployed sign-detection pipeline on Raspberry Pi 5.
Recognizing signs does not amount to general ASL translation. Dataset coverage and differences in signing styles remain important limits. A model walkthrough and selected source excerpts are available on request.
What’s next
Possible directions for further development.
Evaluate a broader range of signing styles and classroom conditions on the target hardware. Further development could improve dataset coverage without treating sign detection as full ASL translation.
Explore the source
Private repository; model walkthrough and selected source excerpts available upon request.