Motion Estimation and Reconstruction for 360° Panoramic Images Based on Variant Goldberg Polyhedral Projection

The effectiveness of transportation systems is significantly influenced by the field of view (FOV) of cameras, with larger FOVs yielding more reliable performance. Conventional cameras with restricted FOV face difficulties in scenarios characterized by field-of-view shake or low-texture imaging areas, complicating feature extraction and matching tracking. Conversely, 360° panoramic cameras, which provide an expansive field of view and high pixel density, are emerging as promising alternatives. While many studies have explored motion calculation based on panoramic vision, most resort to conventional image processing methods for panoramic images, leading to inefficiencies and reduced accuracy. This article proposes a novel approach to panoramic vision motion estimation and reconstruction. The method involves projecting the panoramic image onto a fitting sphere and establishing a mapping relationship between a uniform spherical grid and the pixels of the panoramic image. This enables distortion-free mapping imaging for the entire panorama, addressing issues with matching consecutive frame images caused by distortion in the polar regions. Going beyond conventional epipolar constraints, this paper introduces a geometric constraint of epipolarity within a three-dimensional sphere and derives pose-solving equations tailored for panoramic vision, enabling motion estimation between frames. Additionally, it derives the polar arc equation of panoramic images, accelerating three-dimensional reconstruction of dense point clouds. The sliding window estimation method is used for iterative algorithm optimization, along with an optimization mechanism aligned with the panoramic camera model. Validation on public datasets and hardware experimental platforms demonstrates the proposed method’s significantly improved accuracy compared to existing optimal algorithms.