Interactive and Quantitative Knowledge-Discovery in Large-Scale 3D Tracking Data
Schott, Benjamin · Repository KITopen (Karlsruhe Institute of Technology) · 2018
New developments in tracking technologies combined with enhanced data storage techniques provide powerful ways to collect a tremendous amount of highly resolved object localization data represented as trajectories. Due to the enormous size and complexity of the routinely produced datasets, the systematic analysis and the extraction of relevant knowledge out of the data is a challenging problem. Furthermore, available prior knowledge could be utilized for sophisticated analysis can in many cases not be sufficiently considered for such complex datasets, due to the technical limitations of existing analysis approaches and software tools. Moreover, existing state-of-the-art tracking algorithms are not able to create error-free tracks in the presence of highly dense and noisy data measurements leading to crucial problems coping with the fragmented tracking data. The major contributions of the present thesis are a new concept to systematically incorporate prior knowledge in the knowledge discovery process of large-scale tracking data combining interactive visual exploration and automated trajectory analysis methods. In addition, a new approach to incorporate fragmented tracking data in the analysis of large-scale 3D+t tracking data was developed. Based on the new visualization framework for the interactive handling of 3D+t tracking data, a new approach is presented to transfer complete analysis pipelines to similar datasets. The overall variety of developed methods was consistently validated on newly developed simulated benchmarks allowing to quantitatively investigate the applicability to different tracking databases. In addition, all developed methods were implemented as a platform independent open-source software framework to make the algorithms accessible to the community. Moreover, the successful application of the newly developed methods to large-scale 3D+t tracking datasets of fluorescent light-sheet microscopy images was shown. In particular, the proposed framework was used to separate, quantify and compare cell groups in whole embryo data sets during gastrulation events in early zebrafish development.