Orthophoto-Based Camera Localization in 5-DoF

Mihajlovska, Ina (author), Skočaj, Danijel (mentor), Dobrevski, Matej (comentor)

This thesis investigates five-degree-of-freedom camera localization relative to a georeferenced digital orthophoto (DOP). Scale-Invariant Feature Transform (SIFT) keypoints are detected in the perspective UAV image and the DOP. Their descriptors are compared by Euclidean distance and filtered with Lowe's ratio test. Each matched DOP pixel is converted to horizontal metric coordinates (X, Y) using the raster scale. Because surface elevation is unavailable, all map points are approximated by a common reference plane Z = 0. Within RANSAC, each homography hypothesis is computed from six randomly selected keypoint correspondences and scored by reprojection error over all retained matches. The winning homography is re-estimated from all inliers and decomposed with the calibrated intrinsic matrix to obtain a plane-relative camera transform. Only [Xc, Yc, roll, pitch, yaw] is reported. Across 60 selected regular-domain samples from the OrthoLoC dataset, the mean absolute error (MAE) is 4.13 m in x, 5.40 m in y, 2.12° in roll, 2.58° in pitch, and 0.16° in yaw. The results show that DOP-only localization is feasible, while the larger horizontal, roll, and pitch errors expose the limitations of replacing real surface relief with a single plane.

• Language: English
• Keywords: UAV localization, digital orthophoto, SIFT, RANSAC, homography, camera pose
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187863
• Publication date in RUL: 15.09.2026

Secondary language

• Language: Slovenian
• Title: Lokalizacija kamere na osnovi ortofoto slik v petih prostostnih stopnjah

Abstract

V diplomskem delu je obravnavana lokalizacija kamere v petih prostostnih stopnjah glede na georeferencirane digitalne ortofoto slike (DOP). Značilne točke SIFT se poiščejo v perspektivnem posnetku letalnika in v DOP. Opisniki so primerjani z evklidsko razdaljo ter filtrirani z Lowejevim testom razmerja. Koordinate ujemanj v DOP se s prostorsko ločljivostjo pretvorijo v metrične horizontalne koordinate (X, Y). Ker višine površja niso znane, metoda predpostavi skupno referenčno ravnino Z = 0. RANSAC oceni homografijo, ki se izboljša z vsemi notranjimi točkami in razstavi z uporabo intrinzične matrike kamere. Metoda izračuna položaj kamere v dveh prostostnih stopnjah ter tri rotacijske kote. Na 60 izbranih primerih iz podatkovne množice OrthoLoC so bile dosežene povprečne absolutne napake 4,13 m za x, 5,40 m za y, 2,12° za nagib okoli vzdolžne osi, 2,58° za nagib okoli prečne osi in 0,16° za zasuk. Rezultati pokažejo izvedljivost lokalizacije samo na osnovi ortofoto slik, večje napake pa razkrivajo omejitev predpostavke ravnega površja.

• Keywords: lokalizacija brezpilotnega letalnika, digitalne ortofoto slike, SIFT, RANSAC, homografija, lega kamere