Hand Gesture Recognition Based on Deep Learning

This research paper focuses on the aspect of the interaction between human-computer and recognition of gestures. The paper explores three methods of contactless gesture recognition - vision-based, sound wave-based, and radio frequency signal-based - and highlights the vision-based method. However, the vision-based method is susceptible to interference from various factors, such as different palm sizes, skin colors, object occlusion, and self-occlusion. To address these issues, the paper proposes the use of deep neural networks for gesture recognition. The paper uses MediaPipe to detect 21 landmarks of hands and output their three-dimensional coordinates. The paper also explores four types of deep neural networks such as RNN (Recurrent Neural Network), CNN (Convolutional Neural Network), hybrid (CNN and RNN), and Transformer Encoder - for identifying ten types of gestures (0 to 10). After optimization, the RNN model is selected, achieving an accuracy rate of 99.28%. Based on the detection of hand gestures, we can develop many practical applications of contactless interaction in the future.