RISE-VIO: Robust Initialization and Targeted Pose Robustification for INS-Centric Visual–Inertial Odometry Under Degraded Visual Conditions

Feature-based visual–inertial odometry (VIO) often suffers from initialization failures and tracking drift under degraded visual conditions, such as low-texture regions, abrupt illumination changes, and scenes with a high ratio of dynamic correspondences. We present RISE-VIO, a real-time inertial-navigation-system-centric (INS-centric) visual–inertial odometry system that improves robustness by introducing GNC-style robustification into two failure-critical stages: initialization and per-frame pose estimation. For robust initialization, we develop a GNC-based decoupled rotation–translation initialization module with a two-stage observability gate, consisting of (i) rotation-compensated parallax-rate screening and (ii) a spectral-stability test on the linear global translation (LiGT) system. For online robustness, we design an IMU-prior-guided GNC-EPnP module to selectively downweight or reject outlier correspondences during pose estimation. Experiments on public benchmark datasets show that RISE-VIO achieves more reliable initialization and more stable trajectory estimation in challenging visual conditions while maintaining real-time performance. Additional Monte Carlo perspective-n-point (PnP) evaluations further support the robustness of the proposed pose estimation module under severe outlier contamination.