SM-VINS: A Fast and Decoupled Monocular Visual-Inertial Sensors SLAM System With Stepwise Marginalization

This article investigates the problem of strong coupling between states in the sliding window leading to high computational effort in monocular visual-inertial sensor navigation system (VINS-Mono). We propose a stepwise marginalized VINS (SM-VINS) together with bidirectional optical flow tracking, achieving higher accuracy and lower computational burden. Compared with the monocular camera sensor-based conventional optical flow method in the front-end, the proposed bidirectional tracking offers more accurate positioning, benefiting from forward and backward information jointly. Moreover, a well-designed two-stage marginalization is implemented in the back-end sliding window rather than done all at once. The two stages marginalize the landmarks and the camera pose, respectively, which can effectively eliminate the impact of strong coupling of some states on the system efficiency. The proposed SM-VINS is validated in multiple simulation experiments based on the widely used datasets, EuRoC and TUM VI, while the open-source VINS-Mono serves as its counterpart. The results demonstrate superior positioning accuracy and computational efficiency compared with VINS-Mono, which showcases the potential for practical applications with high accuracy and real-time requirements.