Online object-level SLAM with dual bundle adjustment

Jiaqi Liu, Yongbin Gao, Xiaoyan Jiang, Zhijun Fang

Abstract: Object-level landmarks enable the SLAM system to construct robust object-keyframe constraints of bundle adjustment and improve the pose estimation performance. In this paper, we present a real-time online object-level SLAM. The dual Bundle Adjustment (BA) optimization method, including high and low frequencies, is proposed to optimize the estimated pose. The High-frequency BA (HBA) module is used to quickly estimate the camera pose by matching landmarks of keyframes and feature points of the current frame. Then, the estimated camera pose is used in the Low-frequency BA (LBA) module to improve the trajectory accuracy. The LBA module integrates the object-level landmarks into the pose graph to optimize the camera pose of local mapping. Moreover, we build an additional object detection thread to extract object 2D bounding boxes online. While this paper improves the data association through the depth projection of point-line features and the Euclidean distance of object centroid. Experimental results show that our proposed algorithm effectively reduce the drift error of camera pose estimation and improve the accuracy by a large margin on different datasets.