Multiple-model Based Verification of Japanese Road Data

This paper describes a semi-automatic system for road verification from high resolution orthophotos in an urban context. The system combines several road detection and road verification approaches from current literature to form a more general solution. Each road detection / verification approach is realized as an independent module representing a unique road model and thus a unique strategy. The object-wise verification result of each module is formulated as a binary decision between the classes “correct road” and “incorrect road”. These individual decisions are combined by Dempster-Shafer fusion, which provides tools for dealing with uncertain and incomplete knowledge about the statistical properties of the data. For each road detection / verification module a function for confidence is introduced that reflects degree of correspondence of an actual test situation with an optimal situation according to the underlying road model of that module. Experimental results achieved with four different test sites in Japan demonstrate the potential and confirm the reliability of the new system.