Scene structure registration for localization and mapping

Image registration, and more generally scene registration, needs to be solved in mobile robotics for a number of tasks including localization, mapping, object recognition, visual odometry and loop-closure. This paper presents a flexible strategy to register scenes based on its planar structure, which can be used with different sensors that acquire 3D data like LIDAR, time-of-flight cameras, RGB-D sensors and stereo vision. The proposed strategy is based on the segmentation of the planar surfaces from the scene, and its representation using a graph which stores the geometric relationships between neighbouring planar patches. Uncertainty information from the planar patches is exploited in a hierarchical fashion to improve both the robustness and the efficiency of registration. Quick registration is achieved in indoor structured scenarios, offering advantages like a compact representation, and flexibility to adapt to different environments and sensors. Our method is validated with different sensors: a hand-held RGB-D camera and an omnidirectional RGB-D sensor; and for different applications: from visual-range odometry to loop closure and SLAM.