This paper investigates real-time road detection for an unmanned ground vehicle (UGV) to navigate in a campus environment. A novel vision system with two monocular cameras at different heights and angles is utilized to accomplish both road region detection and direction estimation tasks simultaneously. An Adaboost-based classifier is used in the road region detection in order to handle the road surface's diversity, and a vanishing point tracking technique is used to estimate the road direction. To describe the unstructured road spaces with an accurate model, a RANSAC Spline Fitting algorithm is adopted to delineate the road borders based on the data fusion results from vision and lasers. Extensive experiments are carried out by using a real UGV platform in a campus environment and results show the feasibility and performance of the proposed approach.