DR-SLAM: drift rejection SLAM with Manhattan regularity for indoor environments

Abstract

In this paper, a drift rejection SLAM (Simultaneous Localization and Mapping) method is proposed targeting indoor scenarios, where SLAM generates large drifts due to the lack of reliable features. To provide sufficient features, we leverage multiple feature primitives and geometric restraints (parallel or perpendicular) restraint in man-made environments. Under some satisfy Manhattan World (MW) assumption scene, such as corridors, we can get absolute and drift-free rotation estimation using a Gaussian sphere. By fully utilizing drift-free rotation estimation under MW assumption and the local stability of purely track restricted by point, line, and plane features, our drift rejection SLAM method becomes more accurate and robust. Additionally, by exploiting the constraint of planar motion on ground robot, we propose an ingenious strategy to reduce translation drift by eliminating vertical movement in the Manhattan world. Advantages of our method over other state-of-the-art algorithms are validated on public datasets and real-world experiments. The code is released at https://github.com/WangWen-Believer/DR-SLAM.