Embedded image processing and analysis close to the image sensor is gaining more and more momentum since it allows to reduce required communication bandwidth to the next layer of the processing hierarchy. This saves energy and enables significant efficiency improvement for different use cases. Image processing on a smart image sensor not only comes with benefits but also includes technical challenges.
The focus of this thesis is on the automatic generation of training data for neural networks for pupil detection. Therefore, a 3D-Face Model for eye tracking should be designed with tunable parameters to create synthetic training data. The tunable parameters should include, the camera, the eyes, skin tone, texture of the iris, eye lids, pupil size, different number of glints, reflections and blur. The application should create images with labels of eye properties (e.g. gaze vector, pupil center, glint center positions...).