Robust road marking extraction in urban environments using stereo images

Most road marking detection systems use image processing to extract potential marking elements in their first stage. Hence, the performances of extraction algorithms clearly impact the result of the whole process. In this paper, we address the problem of extracting road markings in high resolution environment images taken by inspection vehicles in a urban context. This situation is challenging since large special markings, such as crosswalks, zebras or pictographs must be detected as well as lane markings. Moreover, urban images feature many white elements that might lure the extraction process. In prior work an efficient extraction process, called Median Local Threshold algorithm, was proposed that can handle all kinds of road markings. This extraction algorithm is here improved and compared to other extraction algorithms. An experimental study performed on a database of images with ground-truth shows that the stereovision strategy reduces the number of false alarms without significant loss of true detection.