Evaluation of Neural Processing Edge Computing Devices for trajectory reconstruction on an IMU-based HCI device
There are numerous situations in daily life where we take notes and afterwards realize that we need the content in a digital form (like meeting notes, protocols, forms or lecture notes). But how to digitize your handwritten notes? There are different products like document scanners, tablets with a stylus or smart pens. All of these solutions require additional equipment or special paper to work correctly. With this thesis we will explore the possibilities of improving a system that is able to digitize handwriting in form of trajectory with a pen that writes on regular paper and which is equipped only with inertial and geomagnetic sensors. Within this work there will be a strong cooperation with STABILO International GmbH.
The primary goal of this thesis is the exploration of possibilities for the adaption and optimization of pretrained neural networks for trajectory reconstruction from inertial and geomagnetic sensor data for a resource constrained embedded device. A pretrained model will be provided by STABILO as well as training data for possible changes in the model after optimization. In the scope of this work the possibilities of Tensor Flow Lite for this task should be explored and optimized models will be evaluated on a number of physical platforms, ranging from general-purpose microcontrollers to specialized SoCs with AI accelerators. Based on the results of the evaluation, a tradeoff between performance and power consumption should be analyzed in detail.
- Basics of Artificial Intelligence, in particular machine-learning methods
- First experiences with Python and C/C++ (embedded)
- Motivation and interest in solving technical problems independently