Title:LF-PGVIO: A Visual-Inertial-Odometry Framework for Large Field-of-View Cameras using Points and Geodesic Segments

Abstract:In this paper, we propose LF-PGVIO, a Visual-Inertial-Odometry (VIO) framework for large Field-of-View (FoV) cameras with a negative plane using points and geodesic segments. Notoriously, when the FoV of a panoramic camera reaches the negative half-plane, the image cannot be unfolded into a single pinhole image. Moreover, if a traditional straight-line detection method is directly applied to the original panoramic image, it cannot be normally used due to the large distortions in the panoramas and remains under-explored in the literature. To address these challenges, we put forward LF-PGVIO, which can provide line constraints for cameras with large FoV, even for cameras with negative-plane FoV, and directly extract omnidirectional curve segments from the raw omnidirectional image. We propose an Omnidirectional Curve Segment Detection (OCSD) method combined with a camera model which is applicable to images with large distortions, such as panoramic annular images, fisheye images, and various panoramic images. Each point on the image is projected onto the sphere, and the detected omnidirectional curve segments in the image named geodesic segments must satisfy the criterion of being a geodesic segment on the unit sphere. The detected geodesic segment is sliced into multiple straight-line segments according to the radian of the geodesic, and descriptors are extracted separately and recombined to obtain new descriptors. Based on descriptor matching, we obtain the constraint relationship of the 3D line segments between multiple frames. In our VIO system, we use sliding window optimization using point feature residuals, line feature residuals, and IMU residuals. Our evaluation of the proposed system on public datasets demonstrates that LF-PGVIO outperforms state-of-the-art methods in terms of accuracy and robustness. Code will be open-sourced at this https URL.