Research on bundle adjustment for visual SLAM under large-scale scene

Abstract

As the scene becomes larger, the cumulative error of Simultaneous Localization And Mapping (SLAM) will increase. In order to reduce the error, the back-end of slam system includes Local Bundle Adjustment (LBA) and Global Bundle Adjustment (GBA). LBA is mainly to optimize the pose of key-frames and landmarks which are in local map, thereby reducing the cumulative error. On the other hand, GBA will optimize all pose of cameras and landmarks in map. However, the number of cameras and landmarks is particularly high, so there are bottlenecks in both memory and efficiency. As to LBA, this paper proposes a new method of screening key-frames and map points, so that the number of key-frames and map points will be as few as possible, while the accuracy of the slam system is high. On the other hand, for GBA, this paper presents a new segmentation strategy for the segmented BA optimization methods to improve the efficiency of optimization and reduce memory consumption.