Enhanced Online Calibration and Initialization of Visual-Inertial SLAM System Leveraging the Structure Information
Dayu Yan, Tuan Li, Chuang Shi
Abstract
Calibration and initialization of point feature-based SLAM systems are prone to performance degradation in challenging environments where rapid motion and texture-less areas are present. Inaccurate calibration and initialization results can introduce additional errors to backend optimization, thereby affecting the localization performance. Difficulty in convergence also limits the practical application of SLAM systems. The structure line segment features present more stability and compact spatial constraint, which can provide additional structural constraints for visual-inertial state estimation. Thus, in this paper, we propose to incorporate the structure information of extracted line segment features and construct an enhanced online calibration and initialization framework for visual-inertial SLAM systems. The proposed method estimates initial velocity, gravity orientation, gyroscope bias, scale, extrinsic rotation, and the potential time delay between IMU and camera data sequence. A vertical structure direction-aided closed-form solution is proposed, achieving both higher initialization accuracy and less convergence time. Experimental results on the synthetic data sequences and real-world dataset prove that our method can achieve robust and accurate calibration and initialization with the introduction of structure information, which is superior to the state-of-the-art works in a high success rate of up to 85% and less convergence time of initialization to 0.5 s within 15 keyframes.