Review of Visual Simultaneous Localization and Mapping Based on Deep Learning

Due to the limitations of LiDAR, such as its high cost, short service life and massive volume, visual sensors with their lightweight low cost are attracting more attention becoming a research hotspot. As hardware computation power deep learning develop by leaps bounds, new methods ideas for dealing simultaneous localization mapping (VSLAM) problems have emerged. This paper systematically reviews VSLAM based on learning. We briefly review development process introduce fundamental principles framework. Then, we focus integration from three aspects: odometry (VO), loop closure detection, mapping. summarize analyze contribution weakness each algorithm in detail. In addition, also provide summary widely used datasets evaluation metrics. Finally, discuss open future directions combining