Real-Time Optimization-Based Dense Mapping System of RGBD-Inertial Odometry

Simultaneous localization and mapping (SLAM) is crucial for intelligent robot applications, especially for mobile vehicles that require accuracy of pose estimation and dense map for navigation. In this paper, we propose a novel SLAM framework based on VINS-Mono, which integrates RGB-D camera and IMU data, can perform high accuracy and robust pose estimation, and construct a dense global map at the same time. A feature tracking method fused with ORB descriptor and multi-view geometric constraints is utilized to improve the features matching accuracy and outlier removal. Depth measurements are used in back-end tightly-coupled based optimization to improve the accuracy of pose and feature depth estimation in the sliding window. Finally, the octree structure and the statistical analysis method are applied to construct the map and make the map clear and consistent. We test our system on public OpenLORIS datasets. The results show that our system has higher pose accuracy and robustness performance compared with state-of-the-art SLAM algorithms, and the map construction is accurate and clear at the same time.