A system for use as a machine guide dog composed of an autonomous vehicle and a two-mirror omni-camera for navigations on sidewalks to guide blind people is proposed. Methods for extracting 3D information from acquired omni-images to localize the vehicle using landmarks of curb lines, tree trunks, stop lines on roads, lawn corners, signboards, and traffic cones are proposed. The methods are based on a space-mapping scheme and three new space line detection techniques. Each space line detection technique can be applied directly on omni-images to compute the 3D locations of a specific type of space line in the landmark shapes. Good experimental results show the feasibility of the proposed system.