PO-AGNC: A Pose-Only-Based Adaptive Graduated Nonconvexity Factor Graph Optimization for Visual-Inertial SLAM

In visual-inertial simultaneous localization and mapping (VI-SLAM), visual residuals are typically formulated using multiview geometry, parameterizing both camera poses and scene feature points as optimization variables. However, as the number of feature points increases, the computational complexity of the optimization problem grows, and such formulations are prone to linearization errors associated with feature points. Recent methods based on pose-only (PO) imaging geometry address these challenges, but the resulting objective function, which involves camera poses at three time instances, is highly sensitive to outliers. To overcome these limitations, we propose PO-adaptive graduated nonconvexity (PO-AGNC), an AGNC graph optimization VI-SLAM system built upon established PO representations. By implicitly representing 3-D points using two base frames, PO-AGNC constructs PO visual residuals to reduce computational complexity and eliminate linearization errors associated with feature points. To handle outliers effectively, we incorporate the Geman–McClure (GM) robust kernel function and employ graduated nonconvexity (GNC) technology. Notably, we propose simultaneously and adaptively adjusting both the control and scale parameters to govern the kernel function’s shape and nonconvexity, steering the optimization toward globally optimal solutions with high efficiency. Extensive experiments on public datasets demonstrate that PO-AGNC outperforms the state-of-the-art methods in terms of efficiency, accuracy, and robustness.