Résumé
Hand gesture recognition has become an essential field in computer vision, enabling natural and intuitive communication between humans and machines. In this project, we developed a lightweight system capable of recognizing static hand gestures in real time from a webcam feed. The method relies on MediaPipe Hands for landmark extraction and a fully connected neural network trained on normalized three-dimensional coordinates of 21 keypoints. Our approach emphasizes efficiency and accessibility by reducing data dimensionality and avoiding the need for complex image-based models. The resulting system achieves high accuracy for multiple gesture classes and can be applied to educational or assistive technologies such as sign language learning tools.