SVM-based road verification with partly non-representative training data

In this paper we present a SVM-based method for automatic quality control of a road database in urban areas. The road verification is carried out by comparing the database objects to high-resolution aerial imagery. The method is trimmed to produce reliable results even if the training data selection is partly non-epresentative. A reliability metric is assigned to the SVM decision that is based on the distance of a test object to the training data. This metric can be applied to any SVM-based classification task. Our experiments show that the classifier is very reliable in only accepting road objects that are actually correct.