Monocular Vision Based Outdoor Scene Understanding For Unmanned Ground Vehicles

Outdoor scenes understanding for Unmanned Ground Vehicles is the core issue of mobile robot adaptation to the environment. Good capability of environment understanding is essential for Unmanned Ground Vehicles’autonomous navigation. In recent years, with the fast development of computer vision, vision-based outdoor scene understanding has become more and more important. Monocular vision-based outdoor scene understanding is more challenging because there are no lane markings or clear edges of the natural scene.In this paper, a combining texture and color method is proposed to detect the road region. Texture orientations are computed using Gabor filters and evaluate its confidence. Then, a local adaptive soft voting (LASV) scheme is used to detect the vanishing point of the road. Detect road boundaries with vanishing-point-constrained edge detection technique. Meanwhile the proposed fast scene segmentation can segment an image into road and non-road regions roughly, which can eliminate the inherent noises in the trees and buildings to improve the robustness. We also use Canny to decrease the voting points to reduce computational complexity.For more complex environment, we present a Conditional Random Fields (CRF) framework with the contextual information of the scene for multi-label classification. Our method use meanshift to segment the images and calculate the color, texton and texture feature. The higher order potential functions used in our framework take the form of the Robust Pn model are more general. DLUT dataset from self-developed UGV platform, Sowerby dataset and CamVid dataset from Cambridge are applied to test our algorithms, experimental results and data analysis show the effectiveness and practicability for various types of environment.