Road detection via unsupervised feature learning

Computer vision based road detection is an indispensable and challenging task in many real-world applications such as obstacle detection in autonomous driving. Low-level image features (e.g., color and texture) and pre-trained models are commonly used for this task. In this paper, we propose a simple yet effective approach to detect roads from a single image, which avoids the supervised model training that typically relies on a considerable number of labeled images. The key idea is to leverage unsupervised feature learning to obtain good road representations. Specifically, we first represent an input road image as a set of image patches. The K-means clustering algorithm is then applied to these image patches (after pre-processing) to generate K cluster centroids. Thus obtained centroids will be used together with a nonlinear mapping function and a bag-of-words projection to derive the image's feature representation in pixel and region levels respectively. All pixels (of the input image) using the former mapping will be clustered by Density Peaks algorithm into several regions, and the regions represented by the latter feature will be grouped by a graph cut method into two classes: road and non-road. Experimental results on several complicated road images demonstrate the effectiveness of our proposed method.