Automatic Road Detection in Traffic Videos

Abstract

Automatic road detection is a challenging and representative computer vision problem due to a wide range of illumination variations and weather conditions in real traffic. This paper presents a novel real-time road detection method that is able to accurately and robustly extract the road region in real traffic videos under adverse illumination and weather conditions. Specifically, the innovative global foreground modeling (GFM) method is first applied to robustly model the ever-changing background in the traffic as well as to accurately detect the regions of the moving objects, namely the vehicles on the road. Note that the regions of the moving vehicles are reasonably assumed to be the road regions, which are then utilized to generate in total seven probability maps. In particular, four of these maps are derived using the color values in the RGB and HSV color spaces. Two additional probability maps are calculated from the two normalized histograms corresponding to the road and the non-road pixels in the RGB and grayscale color spaces, respectively. The last probability map is computed from the edges detected by the Canny edge detector and the regions located by the flood-fill algorithm. Finally, a novel automatic road detection method, which integrates these seven masks based on their probability values, defines a final probability mask for accurate and robust road detection in video.