Abstract
Abstract
Performant and efficient detection and decoding of visual fiducial markers is an im- portant task for tracking unique objects over long periods of time. Having a highly precise and time-efficient method for this task is especially important in context of automated monitoring, data recording and analysis of scientific experiments. In this thesis we propose a performant, efficient and flexible method of detecting and decod- ing fiducial markers in images and video frames. This method is structured as a two- stage marker detection and decoding pipeline made up of independent deep learning models that can be trained separately. We also describe and analyse a method for gen- erating highly effective training datasets of varying size using a small set of source object and background images. Finally we demonstrate the effectiveness of our meth- ods for fiducial detection and decoding tasks using real biological experiment data on Apis mellifera tagged with BEEtag markers.