dukejohannesA/Hand-tracking---NSL-sign-mini-trainer

A browser-based mini-trainer for recognising selected Namibian Sign Language (NSL) fingerspelling letters from a webcam. The application runs hand tracking and classification locally in the browser, so camera frames are not uploaded to a server by this project. # NSL Fingerspelling Mini-Trainer A browser-based mini-trainer for recognising selected Namibian Sign Language (NSL) fingerspelling letters from a webcam. The application runs hand tracking and classification locally in the browser, so camera frames are not uploaded to a server by this project. ## Features - Real-time webcam hand tracking with MediaPipe Hand Landmarker. - Recognition of nine trained letters: `B`, `C`, `D`, `I`, `L`, `O`, `V`, `W`, and `Y`. - Live predicted letter, confidence score, confidence bar, and top-three alternatives. - Reference cards for the supported letters. - Separate data-collection page for recording new landmark samples. - Browser persistence for recorded samples using `localStorage`. - JSON export for retraining the classifier. ## Quick Start This is a static web project and has no `npm` or Python package dependencies for running the trainer. It does need to be served from a local web server; opening `index.html` directly with a `file://` URL can prevent module imports, model loading, or webcam access from working. From the project directory, run one of the following: ```powershell # Python python -m http.server 8000 ``` Or use any equivalent static server, such as the VS Code Live Server extension. Then open: ```text localhost ``` Allow camera access when the browser asks. A modern browser with WebGL support is recommended because the application requests the MediaPipe GPU delegate. ## Using the Trainer 1. Start the local server and open `index.html`. 2. Allow webcam access. 3. Place one hand in view of the camera. 4. Hold the sign steady and watch the prediction and confidence. 5. A prediction is displayed as recognised only when confidence is at least `70%`. The displayed result is smoothed over five frames to reduce flicker. The page mirrors the webcam image and draws the detected hand landmarks over the video. ## Recording Training Data Open `record.html` from the same local server: ```text …