Large-Scale Dense Mapping System Based on Visual-Inertial Odometry and Densely Connected U-Net

Abstract

SLAM-based 3D reconstruction system mainly relies on visual odometry, which uses detailed scanning to obtain better reconstruction results. Among the feature-based SLAM methods, it is difficult to have good reconstruction results where features are missing even when the camera moves slowly. For the direct methods, when exposure changes and blurring occurs, tracking loss will cause unsatisfactory reconstruction effects. This paper proposes a large-scale mapping system based on visual-inertial odometry to solve these problems. The combination of vision and IMU is used to constrain the trajectory estimation in areas where the corners are missing. The system supports depth both obtained by the depth camera and estimated by the neural network. According to the voxel hash mechanism, we only focus on the voxels within the cutoff distance and use the hash table to represent the voxels sparsely to reduce the memory usage. Experiments show that the proposed system can obtain an ideal 3D reconstruction model.