RF-MSCKF: Robust Features-Aided Multi-State Constrained Kalman Filtering for Monocular Visual-Inertial Odometry

Both filtering-based and optimization-based visual- inertial odometry (VIO) methods rely on extracting and tracking feature points across consecutive images. In the context of the Multi-State Constrained Kalman Filter (MSCKF), these feature points are triangulated using their image tracks and the corresponding camera poses in the state vector to perform measurement updates. Consequently, the accuracy of feature extraction, tracking, and outlier rejection directly influences the precision of feature triangulation - and, ultimately, the overall navigation accuracy. However, not all feature points can be reliably tracked across frames, which degrades navigation performance. Therefore, it is essential to extract and track robust feature points that can maintain consistent accuracy, thereby reducing navigation errors and improving trajectory estimation. A further limitation arises from traditional feature extraction methods, such as FAST, commonly used in existing VIO systems. These methods select features solely based on a response threshold, which often leads to the omission of potentially robust features. To address the navigation errors caused by feature extraction and tracking uncertainties, we propose a novel method based on YOLO to identify robust neighborhoods where feature points can be extracted and tracked with bounded errors. This approach enhances the overall robustness and accuracy of monocular visual-inertial odometry.