Optimizing a hybrid feature SLAM system for autonomous driving based on feature compensation

A critical problem in the current simultaneous localization and mapping (SLAM) research is the matching estimation of multi-feature fusion, such as low feature matching accuracy; weak robustness; and poor timeliness. Therefore, this paper focuses on using edge features for assisted localization, which makes two important contributions to multi-feature fusion in autonomous driving scenarios. Firstly, based on inheriting the three-thread architecture of ORB-SLAM2, this paper uses point features to filter and determine new keyframes. Then a feature evaluation method based on map information is proposed, aiming to efficiently introduce edge compensation features and optimize the real-time processing speed of multi-feature based SLAM systems. Secondly, to optimize the accuracy of positioning by edge compensation features, this paper starts from two aspects: the description of angles and the description of length(endpoints) to further define and optimize the reprojection error. The algorithm is tested in the KITTI, an autonomous driving scene data set. And we made a quantitative analysis of the influence of optimization results on computing time and positioning accuracy. Compared to ORB-SLAM2, this system improves localization accuracy by an average of 15.1%, while saving computation time by an average of 26.9% more than PL-SLAM. The effectiveness of the proposed method was verified through simulation in a typical dataset and comparison with the state-of-the-art(SOTA) SLAM system.