Real-Time Gesture Recognition Using Convolutional Neural Networks (CNN)

This paper presents a vision-based gesture recognition system trained using a Convolutional Neural Network (CNN). Gesture data is collected via a standard RGB webcam, pre-processed through grayscale conversion, background subtraction, Gaussian blur, and pixel normalisation, and then fed into a three-block CNN for feature extraction and classification. The proposed model is trained with the Adam optimiser (lr = 0.001, batch size = 32) using categorical cross-entropy loss over 50 epochs with early stopping. On a 10-class custom dataset of 5,000 labelled images, the system achieves an overall macro-average accuracy of 95.5%, with real-time inference running at approximately 28 fps on commodity hardware. Comparative evaluation against state-of-the-art vision-based methods confirms competitive performance without requiring specialised sensors.