Visual Inertial SLAM with Deep Learning-enabled Loop Closure for Challenging Scenes

Loop closure is an important module in a visual Simultaneous Localization and Mapping (SLAM) system. This module effectively uses the previous mapping results to eliminate the accumulated errors from the odometry. The bag-of-words (BoW) model that often used in loop detection and re-localization is susceptible to perceptual confusion in scenes where the building structure is highly repetitive or with poor lighting conditions. Thus, it is prone to large localization and mapping errors. In this paper, we use multiple deep neural networks to improve loop detection and re-localization, respectively. The NetVLAD for place recognition is used for robust loop detection, and the SuperPoint and SuperGlue networks are used for accurate feature matching and re-localization. Our loop closure module is applied to two popular visual-inertial SLAM systems. All functions can run in real-time on the GPU and generally imporved the state estimation performance on EuRoC dataset compared to BoW-based approach. Furthermore, it can provide accurate visual localization when a priori map is available. The code is open sourced on ov_hloc.