Viewpoint-invariant soccer pitch registration using geometric and learned features

Abstract

Automatic registration of broadcast soccer images to a standardized field model enables advanced analytics, augmented reality overlays, and precise player tracking. We propose a fully automatic, viewpoint-independent homography estimation pipeline fusing three complementary geometric cues: white field markings (lines and elliptical arcs), grass-band delimitations, and a binary playing-field mask. Detected primitives are first richly labeled — classifying lines as longitudinal or transversal, characterizing grass-tone transitions, and encoding four-quadrant intersection patterns — to reduce correspondence ambiguity. We then generate and prune candidate subsets of primitives, establish plausible matches to model elements via intersection-pattern rules and projective cross-ratio invariants, and systematically evaluate homography hypotheses using bidirectional mask-projection accuracies and mean reprojection error. An experimental evaluation on the LaSoDa benchmark demonstrates that the proposed method achieves highly accurate registrations with ground-truth primitives and robust performance in the fully automatic end-to-end pipeline. Furthermore, comparative experiments with recent state-of-the-art approaches confirm improved precision and robustness across diverse broadcast scenarios.