Robust Visual SLAM systems with Constrained Sub-pixel Corners and High-quality Line Features

Aiming at the accuracy loss caused by integer corner coordinates and repeated detection of the same line features in the point-line visual Simultaneous Localization and Mapping (SLAM) system, this letter proposed an approach that introduced an optimized bilinear interpolation algorithm in point detector, and then used the iterative method to find sub-pixel points. This letter also optimized the line features during processing. Firstly, we used the restricted short-line rejection strategy to remove some short-lines and added line features in multiple directions. Secondly, this letter proposed a method of three strategies to identify and merge repeated lines on pose-graph estimation. We tested the approach on the EuRoC Micro Aerial Vehicle (MAV) datasets and used the Absolute Trajectory Error (ATE) for evaluation. We also tested it in the real world and compared with the other two algorithms to verify its robustness and accuracy. The contribution of this paper is to verify the feasibility of sub-pixel corners in the point line slam system to improve the accuracy of the system, and combined with the optimization of line features to achieve a higher accuracy system.