Please use this identifier to cite or link to this item: https://olympias.lib.uoi.gr/jspui/handle/123456789/40596
Full metadata record
DC FieldValueLanguage
dc.contributor.authorΜπουζιώτας, Δημήτριοςen
dc.date.accessioned2026-11-20T10:46:54Z-
dc.date.available2026-11-20T10:46:54Z-
dc.identifier.urihttps://olympias.lib.uoi.gr/jspui/handle/123456789/40596-
dc.rightsDefault License-
dc.subjectGeolocation estimationen
dc.subjectcomputer visionen
dc.subjectUnmanned Aerial Vehicle (UAV)en
dc.subjectvegetation mappingen
dc.subjecttarget trackingen
dc.subjectgeospatial data processingen
dc.subjectPMV-Geoen
dc.subjectsemantic segmentationen
dc.subjectmulti-object trackingen
dc.subjectsatellite imageryen
dc.titleUAV-based advanced geolocation estimation with integrated computer vision methods.en
dc.typemasterThesisen
heal.typemasterThesisel
heal.type.enMaster thesisen
heal.type.elΜεταπτυχιακή εργασίαel
heal.secondaryTitleΠροσδιορισμός γεωγραφικών συντεταγμένων αντικειμένων από βίντεο drone με χρήση υπολογιστικής όρασηςel
heal.classificationGeolocation estimation and computer vision
heal.dateAvailable2026-11-20T10:47:55Z-
heal.languageenel
heal.accessfreeel
heal.recordProviderΠανεπιστήμιο Ιωαννίνων. Σχολή Πληροφορικής και Τηλεπικοινωνιών. Τμήμα Πληροφορικής και Τηλεπικοινωνιών.el
heal.publicationDate2026-10-22-
heal.abstractTraditional methods to accurately localize objects such as vehicles, pedestrians, or coastal vegetation over large geographic areas often rely on physical hardware implants like GPS tags, which are costly, labor-intensive, and restricted to pre-selected targets. To overcome these limitations, this thesis presents an end-to-end framework that captures high-resolution imagery and video footage from Unmanned Aerial Vehicles (UAVs) and leverages state-of-the-art computer vision architectures to automate target detection, semantic segmentation, multi-object tracking, and geolocation estimation. The system framework integrates onboard UAV sensor telemetry including active gimbal pitch and yaw angles, GNSS positioning, and relative flight altitude to project 2D image-plane coordinates into 3D global geodetic coordinates (WGS-84). This transformation mathematically models camera intrinsic parameters, asymmetric 16:9 sensor cropping principal point displacements, 5-parameter Brown-Conrady lens distortion, 3D gimbal rotation matrices, and altitude-scaled ray-to-ground intersections under a locally flat ground plane baseline (z = 0). The operational validity of the pipeline is evaluated across two distinct real-world applications: first, an automated Probabilistic Multi-View Geolocation (PMV-Geo) framework coupling deep semantic segmentation (YOLO26x-sem) with Gaussian-weighted rasterization across accumulated probability and density matrices to map coastal salt marsh canopies in the Amvrakikos Gulf (GSD ≈ 4.09 cm/px) without topological polyline fragmentation, evaluated via Symmetric Cross-Correlation Evaluation (SCCE) using Zero-Mean Normalized Cross-Correlation (ZNCC) to achieve 2.90 - 3.17 m mean shoreline precision across multi-year temporal shifts; second, a real-time target detection and tracking pipeline integrating edge-optimized YOLO26 architectures with BotSORT tracking for aerial vehicle and pedestrian localization, utilizing a Within-Bounds (WB) geometric filter to eliminate centroid drift induced by partial boundary truncation along frame edges. Experimental evaluations under controlled flight profiles achieve a track-level median Root Mean Square Error (RMSE) down to 0.498 m, a 50% Circular Error Probable (CEP50) of 0.395 m, and a 95% operational bound (CEP95) of 0.698 m, providing a highly accurate, cost-effective, and scalable solution for ecological monitoring, urban surveillance, and search and rescue operations.en
heal.tableOfContents1 Introduction 11 1.1 Geolocation Estimation and Constraints of Geospatial Mapping . . . . . . . . . 12 1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2 Technical Foundations of Computer Vision 16 2.1 The Vanishing Gradient Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 The Black Box Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.2.1 Explainability & Perception . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3 Convolutional Neural Networks (CNN) . . . . . . . . . . . . . . . . . . . . . . . 19 2.4 Vision Transformers (ViT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.5 Comparative Performance of CNNs & ViTs . . . . . . . . . . . . . . . . . . . . . 22 2.6 Architectural Evolution and Structural Innovations of YOLO Models . . . . . . 24 2.6.1 First Generation Modern Architectures: YOLOv4 and Scaled-YOLOv4 . 24 2.6.2 Architectural Redefinition and Re-Parameterization: YOLOv7 . . . . . . 26 2.6.3 Anchor-Free Detection and Gradient Optimizations: YOLOv8 . . . . . . 27 2.6.4 Breaking the Information Bottleneck: YOLOv9 . . . . . . . . . . . . . . 28 2.6.5 Shift to NMS-Free E2E Object Detection: YOLOv10 . . . . . . . . . . . 29 2.6.6 Complete Architectural Re-definition: YOLOv11 . . . . . . . . . . . . . 30 2.6.7 Attention-Centric Spatial Modeling: YOLOv12 . . . . . . . . . . . . . . 30 2.6.8 YOLO26: Edge-First Small-Target Optimized NMS-Free DFL-Free Ar- chitecture with MuSGD Training . . . . . . . . . . . . . . . . . . . . . . 31 2.7 Synthesis of YOLO Architectural Advancements . . . . . . . . . . . . . . . . . . 33 3 Computer Vision in Aerial Imagery 34 3.1 Challenges of Computer Vision in Aerial Imagery . . . . . . . . . . . . . . . . . 34 3.2 Public Domain CV Datasets for Aerial Imagery . . . . . . . . . . . . . . . . . . 36 3.3 Comprehensive Analysis of the VisDrone Dataset . . . . . . . . . . . . . . . . . 38 3.4 Computer Vision Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . 39 3.5 YOLO Model Performance on VisDrone dataset . . . . . . . . . . . . . . . . . . 40 3.6 Architectural Improvements for Small Target Detection . . . . . . . . . . . . . . 41 3.6.1 Global Context through Area Attention and Partial Self Attention . . . . 42 3.6.2 Multi-Scale Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 4 Geolocation Estimation Pipeline 44 4.1 Study UAV and Intrinsic/Extrinsic Parameters . . . . . . . . . . . . . . . . . . 44 4.2 Asymmetric Sensor Cropping and Principal Point Displacement . . . . . . . . . 46 4.3 Pipeline Implementation and Lens Distortion Modeling . . . . . . . . . . . . . . 47 4.3.1 Extrinsic Parameter and Hardware Constraints . . . . . . . . . . . . . . 47 4.4 Projection of 2D Image Plane to Global Coordinate System . . . . . . . . . . . . 49 4.4.1 Established Image Principal Point Coordinates . . . . . . . . . . . . . . . 50 4.4.2 Target Pixel Displacement from Principal Point . . . . . . . . . . . . . . 50 4.4.3 Gimbal Rotational Transformation . . . . . . . . . . . . . . . . . . . . . 51 4.4.4 Ray To Ground Intersection . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.4.5 Solving the Depth Problem . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.4.6 Scaling Meters to Geodetic Degrees . . . . . . . . . . . . . . . . . . . . . 53 5 Experimental Geolocation Applications and Advanced Mapping Techniques 55 5.1 Application I: Probabilistic Multi-View Geolocation of Semantically Segmented Salt Marshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 5.1.1 Ecological Importance of Salt Marshes . . . . . . . . . . . . . . . . . . . 56 5.1.2 Monitoring of Coastal Salt Marshes . . . . . . . . . . . . . . . . . . . . . 57 5.1.3 Monitoring Coastal Marshes through UAV Imagery . . . . . . . . . . . . 57 5.1.4 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 5.1.5 Computer Vision Pipeline Training & Evaluation Dataset . . . . . . . . . 59 5.1.6 Geolocation Pipeline Evaluation Dataset . . . . . . . . . . . . . . . . . . 60 5.1.7 Detection of Coastal Salt Marshes . . . . . . . . . . . . . . . . . . . . . . 61 5.1.8 Probabilistic Multi-View Geolocation using Gaussian-Weighted Rasteri- zation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 5.1.9 PMV-Geo Evaluation Framework . . . . . . . . . . . . . . . . . . . . . . 68 5.1.10 Symmetric Cross-Correlation Evaluation (SCCE) . . . . . . . . . . . . . 69 5.1.11 Structural Symmetry Scoring algorithms . . . . . . . . . . . . . . . . . . 69 5.1.12 Isolating Geolocation Error to Evaluate Symmetry Scoring Techniques . 71 5.1.13 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 5.1.14 Application I: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 75 5.2 Application 2: Geolocation Estimation of Detected and Tracked Targets in Aerial Footage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5.2.1 Core Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5.2.2 Geolocation Evaluation Framework . . . . . . . . . . . . . . . . . . . . . 78 5.2.3 Data Acquisition & Geolocation Evaluation Scenarios . . . . . . . . . . . 79 5.2.4 Training and Detection of Optimized Networks . . . . . . . . . . . . . . 80 5.2.5 Detection and Geolocation of Vehicles and Pedestrians . . . . . . . . . . 80 5.2.6 Geolocation Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . 82 5.2.7 Application II: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.3 Concludion & Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86en
heal.advisorNameΚαρβέλης, Πέτροςen
heal.committeeMemberNameΚαρβέλης, Πέτροςen
heal.committeeMemberNameΣτύλιος, Χρυσόστομοςen
heal.committeeMemberNameΤσούλος, Ιωάννηςen
heal.academicPublisherΠανεπιστήμιο Ιωαννίνων. Σχολή Πληροφορικής και Τηλεπικοινωνιών. Τμήμα Πληροφορικής και Τηλεπικοινωνιώνel
heal.academicPublisherIDuoiel
heal.numberOfPages96el
heal.fullTextAvailabilitytrue-
Appears in Collections:Διατριβές Μεταπτυχιακής Έρευνας (Masters) - ΤΠΤ

Files in This Item:
File Description SizeFormat 
Μ.Ε Μπουζιώτας Δημήτρης (2026).pdf23.48 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.