| Author | Witold Dębski |
| Supervisor | Grzegorz Galiński, PhD |
| Thesis type | Master’s (MSc) thesis |
| Defence date | 2023 |
Hand gesture detection features in many fields of science. Although solutions for implementing gesture detection do exist, most of them require considerable effort, which creates a barrier for individual developers or small teams wishing to use gesture detection in their applications.
The thesis presents the development of the GestRec library, which aims to lower the barrier to entry for deploying gesture detection in applications. Available solutions and datasets were analysed, a custom dataset was prepared and used to train popular neural network models, achieving an accuracy of 95.7% with the ResNet34 model. In addition, a gesture classification solution based on Google’s MediaPipe combined with an SVC algorithm was tested, achieving 93.4%.
Once a suitable network model had been prepared, the GestRec library, which classifies human hand gestures in real time, was created and released publicly. The library comes with automatically generated documentation
and two example applications demonstrating its use, which users can build on in their own projects. A number of optimisation techniques were applied in the library, enabling real-time processing at more than 15 frames per second without the use of the CUDA platform.