A robust road segmentation method based on graph cut with learnable neighboring link weights

Road region detection is a crucial functionality for road following in advanced driver assistance systems (ADAS). To address the problem of environment interference in road segmentation through a monocular vision approach, a novel graph-cut based method is proposed in this paper. The novelty of this proposal is that weights of neighboring links (n-links) in a s-t graph are estimated by Multilayer Perceptrons (MLPs) rather than calculating by the neighboring contrast simply in previous graph-cut based methods. Estimating n-link weights by MLPs reinforces the ability of graph-cut based road segmentation algorithms to tolerate the complex and changeable appearance of road surfaces. Additionally, the Gentle AdaBoost algorithm is integrated into the graph-cut framework to estimate the terminal link (t-link) weights in the s-t graph. Experiments are conducted to show the robustness and efficiency of the proposed method.