Title: Improving The Applicability Of Visual Slam With Submodular Submatrix Selection
Dr. Vela, Advisor
Dr. Yezzi, Chair
The objective of the proposed research is to improve the applicability of Visual Simultaneous Localization and Mapping (VSLAM) on diverse platforms and scenarios, which has broad impact on practical applications in Robotics and Argumented Reality (AR). Recently, the applicability of VSLAM on diverse platforms (e.g. micro flying vehicle, AR headset) and scenarios (e.g. low-texture, fast motion) draws attention in SLAM community. However, state-of-the-art applicable VSLAM involves design choices that trade efficiency with significant performance loss. In this proposal, we revisit feature-based Bundle Adjustment (BA) SLAM, which has high performance but also high computation cost. Our work is closely related to submodular submatrix selection, which is originally studied in large-scale machine learning problem. Throughout this research, the idea of submatrix selection is applied to both front and back ends of feature-based BA VSLAM, with improvements on multi-perspectives: efficiency, robustness and accuracy. Two works are presented as preliminary research. The 1st work, good feature matching, enables efficient frame-to-map feature matching, while preserving the accuracy of pose tracking in feature-based VSLAM. The 2nd work, good line cutting, improves the robustness to line triangulation error in line-assisted VSLAM, therefore improves the applicability on challenging scenarios where point-feature fails. In the proposed work, the submatrix selection idea is extended to factor graph building, which will improve the efficiency of BA with little performance loss.