Continuous Gesture Recognition in Human-Computer Interaction through Real-Time Machine Learning Models

This paper focuses on the realisation of gesture recognition for permanent interaction with the human-computer interface using deep learning. The study focuses on the fact that the necessary gestures are dynamic and continuous, and the approach is based on the architecture that is CNN-LSTM hybrid for spatial features and temporal sequence modeling. The training and evaluation of the system are conducted on gesture datasets which are open source while real-time implementation is done on the edge devices by using the TensorRT. This proves its efficiency and promptness and the performance indicators are provided with regard to various scenarios. Research on the users showed positive responses concerning the unit and the degree of engagement the users had with the system. Nevertheless, the issues of generalization and increasing the computational speed are the main concerns for further development However, the presented work offers promising perspectives for the improvement of the gesture-based interfaces in the applications from virtual reality to assistive technology. In future research, we will improve the architecture proposed in the paper, add new features, make the system open to more kinds of data modalities, and enable it to learn more gestures and contexts for interactions in general.