The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Download
Citation
Articles | Volume XLVIII-1/W1-2023
https://doi.org/10.5194/isprs-archives-XLVIII-1-W1-2023-503-2023
https://doi.org/10.5194/isprs-archives-XLVIII-1-W1-2023-503-2023
25 May 2023
 | 25 May 2023

EFFECTS OF HYPER HEMISPHERICAL FIELD IN BUNDLE ADJUSTMENT WITH FISHEYE IMAGES

A. M. G. Tommaselli, T. A. C. Garcia, L. F. Castanheiro, M. B. Campos, and G. H. Santos

Keywords: Fisheye camera, Hyper Hemispherical lenses, bundle adjustment, Mobile Mapping, Backpack platform

Abstract. The large field of view and compact structure make fisheye lens cameras an attractive technology for mobile mapping systems and visual navigation. Fisheye cameras usually have a field of view equal to or higher than 180°. Lenses that can capture light rays coming from angles larger than 180° are known as hyper hemispherical lenses. Some of the existing mathematical models can be unsuitable for those points beyond the 180° field of view when performing photogrammetric processes based on these equations, underexploring the full potential of fisheye lens cameras. In this case, depending on the selected projection model, points appearing in the hyper hemispherical (HH) field can produce blunders in the bundle adjustment. Nevertheless, most of the available solutions for camera calibration and bundle adjustment were implemented using the equidistant model. Therefore, the points located in the hyper hemispherical field are often removed from the bundle adjustment either during keypoint detection or by applying a mask to remove the entire HH field from the images. In this paper, we assessed experimentally the hypothesis that using the original HH full-field image introduces blunders, deteriorating the bundle adjustment. The experiments were performed with a Ricoh Theta S 360° camera which was mounted on a backpack platform. The camera was set in video mode, and frames were captured at a rate of 1 fps, while traversing an urban street, generating 307 frames. Experiments were performed, including or removing HH points in the bundle adjustment. The results show that the errors in the dataset with the hyper hemispherical field were larger than the ones using cropped images.