Manuscript received December 9, 2025; revised January 19, 2026; accepted April 2, 2026; published September 29, 2026.

Abstract—Fisheye cameras offer images covering a large Field of View (FOV), typically a hemisphere for a single camera, with the trade-off of significant radial distortions and object rotations. The literature indicates that a Frames-Per-Second (FPS) friendly approach creates multiple gnomonic projections, with the aim of covering the complete hemisphere, i.e., fisheye image. These projections are naturally similar to pinhole perspective cameras, thus allowing restoration of the correct object orientation and reduction of the distortions per-projection. Hence, we can make use of the abundant resources related to perspective cameras for fisheye imagery. However, the choice of projection parameters—number of views, their arrangement, field of view, longitude and latitude of the views—has received little systematic study. We show that a hypothesis-driven optimization of these parameters yields moderate to substantial Average Precision (AP) gains over the existing baseline, with AP improvements reaching up to ~+10% across three public benchmarks. Additionally, the existing method backprojecting bounding boxes from the projected views to the original fisheye is altitude-dependent. To suit the drone imagery case, we propose an altitude-agnostic method, and use it for all projection settings we tested. To further confirm our optimized configuration, we conducted extensive tests isolating the effect of projection parameters, You Only Look Once (YOLO) backbones, and input resolutions.

Keywords—fisheye camera, gnomonic projection, object detection, Field of View (FOV), drone imagery
 

Cite: Yassir Zardoua, Joelle Jreis, and Joanna Faddoul, "Pedestrian Detection in Top-view Fisheye Images for Drone Applications," Journal of Image and Graphics, Vol. 14, No. 5, pp. 817-829, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).