GR2AM, is the project we created for the Physical Computing Seminar, at Saarland University with the HCI Lab. GR2AM is a proof of concept towards creating a simple-to-use software in the scope of Human-Computer Interaction, to create a gesture recognition, and application-mapping system, in hopes of facilitating the interaction with smart computing and IoT devices. GR2AM is a hybrid-learning based model, with a user-friendly interface, to easily record gestures, either from a predefined set of nine gestures, or create up to six unique custom gestures, map them to trigger certain applications, and process incoming, real-time, data streams. The hybrid-learning model consists of a 1D-CNN and a Random-Forest Classifier, trained on a minimal dataset of 10 samples for each gesture, to produce a real-time accuracy of 82.85%, from a stream of 30 frames of 21 hand landmarks captured using a built-in webcam by using the MediaPipe Python Module.
- MediaPipe & OpenCV, for gesture capturing
- PyTorch, for the 1D-CNN
- Scikit-learn, for the Random-Forest Classifier
- Web UI
- Flask, for the backend web framework
- Bootstrap, for the frontend framework
- Jinja, a template engine
- CSS, for styling
- Javascript, for all the interactive parts of the UI