Dense Reconstruction from Visual SLAM with Probabilistic Multi-Sequence Merging

Abstract

This thesis presents a comprehensive visual SLAM system that extends the application of ORB-SLAM3. Using it as a template, a supplementary and optional function of 3D dense reconstruction is implemented for both RGB-D and stereo cameras. With conventional datasets, TUM, EuRoC, and KITTI as benchmarks, we confirm the validity of proposed system in both indoor and outdoor scenarios. Besides, the concept of Octree is integrated into our system to generate Octomap. A compact mapping can be achieved as such, verified by the fact that the size of each dense point cloud map is reduced to approximately one-fifth after the conversion. Furthermore, a multi-sequence merging method is included in our proposed system, formulating with a probabilistic-based optimizing algorithm and map accessing functions from the original system. Multi-sequence experiments evince that the tracking accuracy profits from the exploitation of a priori knowledge gathered through the preceding sequences.