Résumé
Fisheye imaging exemplifies this paradigm shift: deployed at scale in autonomous and edge AI systems, these sensors produce data consumed almost exclusively by machines, yet their strong radial distortions remain a largely unsolved challenge for conventional feature extraction pipelines. Feature matching in particular suffers, as standard detectors assume locally planar geometry, an assumption that breaks down in the peripheral regions where fisheye lenses capture the most scene content. The proposed method, SURFisheye, is a radially-aware extension of SURF (Speeded-Up Robust Feature) built on the equidistant fisheye projection model. A position-dependent scale adaptation factor enlarges box filter support toward the image boundary, while a radiallydecreasing Hessian threshold preserves peripheral responses; both corrections are propagated to the Haar wavelet descriptor stage. Requiring no camera calibration or image rectification, SURFisheye outperforms SURF by factors of around 2 times, while maintaining a validity ratio above 98% and preserving twice as many matches under threshold filtering.