Visual Compass Utilizing Structural Regularities for Drone Navigation

Abstract

This thesis propose the San Francisco world (SFW) model, a novel structural model inspired by San Francisco’s hilly terrain. Our SFW consists of a single vertical dominant direction (VDD), two horizontal dominant directions (HDDs), and a slope. SFW is a more general model than the Manhattan world (MW) and a more compact model than the Mixture of Manhattan world (MMW). Leveraging the structural regularity of SFW such as uniform inclination and periodicity of slopes, we design an efficient and quasi-globally optimal DD/VP estimation method through an aggregation of sloping line normals on the Gaussian sphere. We further leverage the structural patterns of SFW for the 3-DoF visual compass, the rotational motion tracking from a single line and plane. Our methods demonstrate enhanced adaptability in more challenging scenes and the highest rotational tracking accuracy compared to state-of-the-art methods. We release the first dataset of sequential RGB-D images in San Francisco world and open source codes.