3D Reconstruction of Indoor Scenes Based on Feature and Graph Optimization

In this paper, we propose an approach of visual SLAM (Simultaneous localization and mapping) system, in which we build 3D scene models and estimate the camera poses at the same time. We detect SURF feature points from RGB images and use their relevant depth data to build 3D correspondences. With these correspondences, we align consecutive frames together using RANSAC, while ICP is performed when RANSAC fails. Keyframes are selected based on visual overlap between current frame and the latest keyframe. Furthermore, we detect loop closures between current keyframe and selected candidates of keyframes. Among these candidates, we perform a radius search to further prune the search space. With both sequential constraints and loop closure constraints, a pose graph is created and optimized using a robust global optimizer based on line processes.