Geotri-VIO: geometrically consistent multi-prism projection for panoramic imaging in visual-inertial-odometry

摘要

Abstract Panoramic camera-based Visual-Inertial-Odometry (VIO) systems play a crucial role in robotic navigation, autonomous driving, and virtual reality applications, owing to their large Field-of-View and enhanced localization capabilities. However, the nonlinear distortions caused by the lack of geometric consistency in projection models for panoramic images pose significant challenges to feature extraction and tracking algorithms. In this paper, we present Geotri-VIO, a novel VIO system that addresses these challenges using a multi-prism projection model. By constructing the multi-prism projection planes such that each face is tangent to the inherent projection sphere of the panoramic camera, the proposed model ensures strict geometric consistency in each projection plane while maintaining global geometric consistency, which is supported by mathematical proof. Additionally, we evaluate the impact of increasing the number of projection planes and demonstrate that triangular prism projection outperforms other multi-prism projection models. To validate its effectiveness, Geotri-VIO is tested on public datasets. Experimental results show that the triangular prism projection significantly improves the tracking accuracy of both point and line features, thereby enhancing the overall localization performance of the VIO system.