Erklärung zur Verfassung der Arbeit

Co-registered intensity and range imaging is one example of complementary multi-modality that has been extensively explored for 3D reconstruction, yet not fully researched for robust estimation. The classic RANdom SAmple Consensus and its variants such as PROSAC, Preemptive RANSAC and SCRAMSAC, to name a few, have been the most commonly applied approaches to visionbased homography estimation, with single modality input as a precondition. In this thesis, RANSACbased homography estimation that builds on the properties of both the 2D intensity modality and 3D spatial modality is to be explored, extended and evaluated with focus on robust estimation. The outcome of this thesis is an extension to the RANSAC algorithm, termed FT-RANSAC, that builds on state-of-the art in RANSAC approaches, and enables a robust homography estimation using both the 2D intensity modality and the 3D point cloud modality inputs. The proposed FT-RANSAC approach has demonstrated its ability to exceed singlemodality state-of-the art in robustness and stability, with implementations as a stand-alone and as an ROS node, and with cross-platform functionality on both Intel x86 and ARM Cortex-A15 architectures. Early contributions of this work have been published in the workshop on Pattern Recognition in Remote Sensing 2014, in conjunction with ICPR 2014, Stockholm, Sweden. This work has been sponsored in-part by Toyota in collaboration with OpenPerceptions Inc.