Segmentation and Vectorization of curbstones from high-resolution ortho images for test sites in Bavaria, Germany

Abstract

Generating road networks manually has always been an ineffective and labour-intensive task. Accurate representation of road networks is crucial for various applications, including urban planning, infrastructure management, navigation systems, and especially autonomous vehicle development. In the realm of autonomous driving, the accurate detection of curbstones holds particular significance, as they serve as critical boundaries for vehicle navigation and safety. However, current methods for online curbstone detection at the site are fraught with challenges, including real-time processing constraints and environmental variability. Fortunately, the availability of high-resolution aerial imagery presents an opportunity for offline curbstone detection, enabling more comprehensive and accurate mapping of road networks. In this thesis, we address the problem of curbstone detection as an iterative graph generation task, wherein curbstone edges are detected vertex by vertex from initial curbstone candidates identified through segmentation. Leveraging techniques from imitation learning, we take a high-resolution ortho-image as input and output a graph representing the detected curbstones. Our approach endeavours to enhance the accuracy and robustness of road edge detection through several enhancements. We introduce a loss function, termed Slope Penalty loss, aimed at refining the model training process by addressing the slight variations in gradients of the predicted vertices. Our experimental evaluations underscore the effectiveness of these enhancements, as demonstrated through comparisons with the already existing curbstone detection algorithms. The proposed approach is tested over the city area of Munich, Bavaria, Germany.