Deep Learning Model for Vision-Based Dynamic Hand Gesture Recognition

Abstract

Communication is essential in our daily life; however, senior citizens may experience communication difficulties due to declining abilities, making it difficult for them to request assistance when needed. The aim of this project is to develop a vision-based hand gesture recognition system to assist senior citizens, where the system will recognise a variety of gestures and translate them into their corresponding meanings for each gesture. A collection of hand gesture videos consisting of 14 gesture classes is used to build the models representing daily tasks. This solution is particularly valuable for bedridden senior citizens who face difficulty communicating with their caretaker. In this study, we compared the performance of three deep learning models: MobileNet + BiGRU, LSTM, and Transformer + DenseNet. Our findings revealed that LSTM outperformed the other two models, achieving a commendable accuracy rate of 94.33%. In comparison, MobileNet + BiGRU achieved an accuracy of 92.20%, and Transformer + DenseNet achieved an accuracy of 87.23%.