Accurate Loop Closure Detection in Semantic SLAM with Projections of Object Models

Loop closure detection is crucial in Simultaneous Localization and Mapping (SLAM), which can correct errors accumulated as mobile robots move in unknown environments. The conventional methods of loop closure detection leveraging Bag of Words(BoW) to detect loop closure are low accuracy and the accumulated errors are significant. Semantic information is deeply integrated into SLAM and provides object-level features to aid loop closure. In this paper, a hierarchical loop closure detection method is proposed, which guarantees the accuracy of loop closure by projections of object models. Objects in the scene are constructed as ellipsoid models by semantic information. In all directions, the projections of object models on the image plane are ellipses. First, two keyframes are detected whether in the same location by comparing the categories and the bounding boxes of semantic information. Second, when two keyframes are considered in the same location, the objects observed by the earlier keyframe of the two keyframes are projected on two keyframes with the form of object models. An error function is established by projections of object models on two keyframes to guarantee the accuracy of loop closure. The proposed method is evaluated and compared to other methods on KITTI dataset, demonstrating the accuracy and robustness.