Robust Stereo Visual-Inertial SLAM Using Nonlinear Optimization

In the case of image blurring, excessive motion and lack of features, the robustness and accuracy of feature-based simultaneous localization and mapping (SLAM) system decline dramatically or even fail. For this problem, a tightly coupled stereo visual-inertial SLAM system using nonlinear optimization is proposed. Firstly, taking the pose of the keyframes as a constraint, the biases of the inertial measurement unit (IMU) are estimated using the divide and conquer strategy. In the front-end, in order to solve the problem that ORB-SLAM2 fails to use the constant velocity motion model due to excessive motion during the tracking process, the initial pose of the current frame is predicted by pre-integrating the IMU data from the previous frame to the current frame, and IMU pre-integration constraints are added to pose optimization. Then, in the back-end optimization, keyframe poses, map points, and IMU pre-integration are optimized within a sliding window and IMU deviations are updated. Finally, the performance of the system is verified with the EuRoC dataset. Comparing with the ORB-SLAM2 system, VINS-Mono system and OKVIS system, the accuracy of the proposed system is improved by 1.14, 1.48 and 4.59 times, respectively. Comparing with the state-of-the-art SLAM systems, the robustness of the system is improved under the conditions of excessive motion, image blurring, and lack of features.