Large-scale Hybrid 3D Map and Line Detection with Uncertainty for Vision-based Self-localization

Towards large-scale environment, a novel metric-topological 3D map is proposed in our vision-based self-localization system. Based on probabilistic line elements with directional information, the local metric map is developed using different feature levels. Then, the adjacent local metric maps are connected by topological structures. We design a nonlinear camera model which propagates directional map elements into the image with uncertainty manipulation. In addition, the associated edge pixels are fitted by solving generalized eigenvalue problem with covariance propagation. A human machine interface is developed for self-localization system. Experimental results reveal that the system can realize map-based line detection with partial occlusion.