Research On Road Segmentation Guided By High-level Structure Information

Scene understanding is one of the key techniques for unmanned ground vehicles. Among the aspects of scene understanding, road segmentation and detection are important components. The unstructured road detection technique has become a hot research topic recently. Compared with the structured road, the unstructured road scenes are usually more complex and capricious. Moreover, the local features of road images(especially the scale of the local features) are different in different locations, due to the perspective effect in the image capturing stage. As a consequence, difficulties are encountered when using traditional feature extraction methods. In order to deal with this problem, the scales of the local road features are estimated combining the precise depth information provided by a Laser ranger. Hence, a multiscale unstructured road detection method is proposed in this paper. Meanwhile, in order to provide a subjective evaluation on the unstructured road detection algorithms, an open data set is established. The main contributions of this paper are as follows:First, a multiscale unstructured road detection method is proposed in this paper. According to the characteristics of the unstructured roads, a road segmentation algorithm combining the color histogram with the gradient histogram is proposed and validated in this paper. Based on this joint feature, a method to generate the road probability maps and detect the roads is introduced using KNN. In order to further take the perspective effect into account, the information of the vanishing point is used to modulate the scale and orientation of the local features. In the proposed method, the local scale and orientation for each location in an image are calculated under the guidance of the depth information provided by the Laser ranger and the high-level information(the information of the vanishing point). Then the color feature is calculated considering the current location and its corresponding scale, and the gradient feature is calculated considering the the current scale and orientation. Experiments show that the performance of road segmentation and detection can be improved using the low-level features considering the scale and orientation.Second, there is no ground truth data set for evaluating the unstructured road detection algorithms in the literature. For this reason, an open data set of unstructured road scene images is established in this paper. In the proposed data set, except for clustering the pixels into road and non-road categories, the concept for uncertain region is introduced for that the boundaries of non-structural roads are usually not distinct. In order to provide a subjective evaluation of the unstructured road detection algorithms, several concrete criteria are founded for the proposed data set. Consequently, the performance of detection algorithms can be compared with each other.