IRP-VINS: A Rotation Prior-Based Visual-Inertial Odometry Method

Abstract To address the problems of feature tracking failure and accumulated trajectory drift in visual-inertial odometry under aggressive rotational motions of unmanned aerial vehicles, this paper proposes an IMU rotation prior-enhanced visual front-end method integrated into the VINS-Fusion framework. The proposed method utilizes the bias-corrected IMU pre-integration states maintained by the backend to construct a feature position prediction model, providing reliable initial estimates for optical flow tracking. This transforms the conventional large-range unconstrained search into a local refinement process guided by motion constraints. Furthermore, a feature selection mechanism combining motion consistency deviation and statistical gating is designed to adaptively identify and remove abnormal observations affected by rapid motion and image blur, thereby improving the quality of measurements provided to the backend optimization. Experimental results on high-dynamic sequences of the EuRoC dataset demonstrate that the proposed method improves feature tracking robustness under rapid rotational motions, effectively reduces trajectory drift, and achieves more stable visual-inertial localization performance in complex motion scenarios.