No abstract

Understanding the road scene structure is essential and important for perceiving the driving situation in intelligent transportation systems (ITS). In this paper, we aim at analyzing the road scene structure by classifying the pixels to three different types, including road surface, lane markings, and nonroad objects. Instead of detecting these three objects separately in traditional approaches, we integrate different ad hoc methods under the conditional random field framework. Three feature functions based on three cues including smoothness, color and lane marking segmentation, are used for pixel classification. Besides, an optimization algorithm using graph cuts is applied to find the solutions efficiently. Experiments on the data sets demonstrate high classification accuracy on objects in the road scene.