Lane Model Validation: Ground Truth Generation and Lane Model Evaluation

Internship Report presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced AnalyticsGenerating ground truth data for training models that are supposed to replace humans for certain tasks, such as in the field of autonomous driving is a big issue for many researchers all over the world. Over different problems in this field there a various approaches to deal with a ground truth generation that does not rely on time consuming and expensive labelling, yet being able to evaluate the performance of models not only qualitatively. Most of the quantitative approaches are using camera images and some are considering GPS data as well. In this report, the data used is the output of a line detection algorithm including positional information per frame and GPS data. Based on the localization of both vehicle and lines, the model can be evaluated by its ability to detect road geometries. The approach results in an estimation of road boundaries that is based on real road markings, but depends on a good parameter choice and input quality. Nevertheless, it is a rather fast and inexpensive way to generate a ground truth that can be compared to the model output in order to evaluate its performance on detecting a valid road geometry