Institutional Repository of Vilnius University
Objekto aptikimas pasitelkiant ypatingųjų taškų radimo SIFT algoritmą /
Abstract
dc:descriptionIn recent years, the rapid growth of unmanned aerial vehicle (drone) technologies has introduced new security challenges across various domains. One of the biggest concerns is the use of drones for smuggling contraband across national borders or flying over restricted areas. Traditional aerial surveillance systems typically rely on radio frequency or acoustic monitoring technologies. However, advancements in drone capabilities have highlighted the growing need for visual detection methods. A literature review identified three primary challenges in visual drone detection: varying drone sizes and distances between the drone and camera, the applicability of the SIFT (ScaleInvariant Feature Transform) algorithm to realtime data, and the complexity of detecting drones against diverse sky backgrounds. To evaluate how well different algorithms address these challenges, a dedicated dataset was created, and a series of experiments were conducted. Experimental analysis revealed that both SIFT and URSIFT algorithms were too slow for realtime applications and demonstrated limited accuracy, particularly when the drone appeared significantly smaller than the reference object or occupied only a small portion of the image. To address the issue of inconsistent performance on drones of varying sizes, a dualbranch model named SIFTRGBCNN was developed and trained. This architecture processes SIFT features in one branch and raw RGB pixel values in the other. Among all models tested, the SIFTRGBCNN achieved the highest drone recognition accuracy in sky imagery, reaching 87.09% and outperforming all other evaluated models. However, it performed the worst in detecting drones smaller than 30x30 pixels, achieving only 25.58% accuracy. Based on the conducted study, it can be concluded that integrating SIFT features with convolutional neural networks in a dualbranch architecture enhances overall detection accuracy but also results in a loss of the SIFT algorithm’s inherent robustness to scale variations and image rotations.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Anužas, Danielius,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD239029804&prefLang=en_US