Pose Correction for SfM-based River Mapping using the Fixed Baseline and Optical Axis of an Omnidirectional Camera as Constraints Teruhiko Meguro (a*), Tetsu Yamaguchi (a), Nobuaki Kubo (b), Etsuro Shimizu (b), Masafumi Nakagawa (a)
a) Shibaura Institute of technology, Japan
*ah20092[at]shibaura-it.ac.jp
b) Tokyo University of Marine Science and Technology, Japan
Abstract
In recent years, the acquisition of high-density point cloud data has been promoted as a way to create digital twins of cities. However, data preparation for urban river spaces has not advanced sufficiently. The two main methods of acquiring 3D data in these environments are LiDAR and image-based 3D measurement. The LiDAR method has the advantage of directly measuring distances and acquiring dense point clouds. However, the accuracy of point cloud acquisition depends heavily on exterior orientation estimation, which requires expensive GNSS/IMU systems. On the other hand, image-based point cloud generation primarily uses structure from motion (SfM) and multi-view stereo (MVS). Although SfM/MVS typically requires more time to generate point clouds compared to LiDAR, it enables the construction of more affordable measurement systems. However, matching large-scale image datasets involves high computational costs. In particular, the linear trajectories required for boat-based measurements along urban riverbanks pose challenges. Long measurement paths can lead to accumulated matching errors and distortions in the point cloud. To address this, we propose a method to improve the robustness of SfM-based pose estimation using omnidirectional images captured from a boat. First, mask images are generated to exclude water surface and occlusions, thereby enhancing image matching reliability. Then, constraints are introduced into the SfM process by leveraging the fixed baseline and optical axis alignment of the omnidirectional camera. These constraints stabilize image pair selection and pose estimation by guiding the image matching process with prior knowledge of camera arrangement. Finally, point clouds are generated using SfM/MVS, with reduced matching errors along the measurement trajectory. We also compared these results with those from LiDAR-SLAM and conventional methods.
