Image-Based Geo-Specific Road Database Creation for Driving Simulation

Geo-specific road databases are sometimes necessary for driving simulation studies. However, manually creating them is very time consuming. One of the reasons is that a real road can have non-uniform cross sections due to irregular shaped road boundaries. Additionally, in the UCF driving simulator system, the road database format requires identification of repeatable cross sections, which is also a labor intensive process. While some commercial software applications, such as the MultiGen Road Tool have road modeling functions, roads with non-uniform cross-sectional profiles are still hard to model. In this research, an image understanding based method is proposed for geo-specific road database development. Digital Line Graph (DLG) data, issued by the United States Geographical Survey (USGS) is used to limit the search space for road segmentation. The mean-shift clustering method is chosen to be the solution for separating pavement and non-pavement areas within road areas. The Linde-Buzo-Gray (LBG) vector quantization method is used to identify repeatable cross sections. The proposed method results in an average accuracy of road segmentation above 85%. The average error of cross-sectional profile is no more than 0.36 feet. The proposed system can significantly reduce the efforts in creating geo-specific road databases. The utilization of the two types of geographical information, high-resolution aerial photos and USGS DLG data, play the most important role in the proposed system.