Road marking detection based on structured learning

Road marking is a key visual cue for driving in structured environments like highways and urban roads. Road marking detection plays an important role in advanced driver assistant systems and autonomous driving. Robust road marking detection is challenging for the variation of road scenes, the degradation of the markings and the changes of the illumination. Traditional algorithms mainly use the grayscale cues or edge cues to detect the markings. However, these approaches are not adaptive to changing environments and the threshold values are hard to select. In this paper, we propose a novel data adaptive structured learning based road marking detection algorithm. A structured random forest is learned to classify each image patch to get a structured label patch. In this way, the contextual information of the images and the structural information of the labels can be effectively exploited to reduce the ambiguity. The channel features including the color name, the normalized gradient magnitude and the histogram of oriented gradient, together with the local self-similarity are extracted as the features of the patches. Experimental results show that the proposed method outperforms the traditional ones.