As part of the thesis, the use of convolutional neural networks on the edge in an interactive exercise system is presented. The work combines the fields of deep learning, software, hardware, and the OAK-1 camera.
The final system guides the user during exercise execution and can be expanded to an arbitrary number of exercise points for longer routines.
The DepthAI library is used, and with its pipeline system, processing is performed directly on the OAK-1 camera, which offloads the host system that only displays the results and graphical interface.
As an implementation example, one exercise and one gesture are included, enabling the execution of a single exercise point. The system is deployed on a final embedded computer, Odroid N2+, and together with a screen, forms a standalone unit.
A comparison of the power consumption of the final system with other implementations confirms that it is energy efficient and suitable for long-term operation.
|
