Robust 3D-SLAM With Tight RGB-D-Inertial Fusion

Abstract

In case of image blur, excessive motion and lack of features, simultaneous localization and mapping(SLAM) based on visual features has the problem of robustness and precision degradation or even failure. A tightly coupled, nonlinear optimization-based RGB-D-Inertial system is proposed for this problem. At the front-end, for the problem that feature-based SLAM fails due to excessive motion during the tracking process, the initial pose of the current frame is predicted by pre-integrating the inertial measurement unit(IMU) data from the previous frame to the current frame. In the back-end optimization, the camera pose and IMU pre-integration are optimized in the sliding window with the pose of keyframe as a constraint. Finally, comparing with the Orbslam2 system, experiments show that the robustness of the RGB-D-Inertial SLAM under the conditions of fast motion, image blur and lack of features is improved. Meanwhile, the accuracy is also enhanced under large-scale environment.