Date of Publication
11-2025
Document Type
Master's Thesis
Degree Name
Master of Science in Applied Physics
Subject Categories
Physics
College
College of Science
Department/Unit
Physics
Thesis Advisor
Edgar A. Vallar
Defense Panel Chair
Maria Cecilia D. Galvez
Defense Panel Member
Jazzie R. Jao
Floro Junior C. Roque
Abstract (English)
Searching for people in distress is quite problematic due to the size and unpredictable nature of the maritime environment, especially as visibility deteriorates the objects become very small against the water surface. Unmanned Aerial Vehicles (UAVs) are used to speed up search and rescue (SAR) operations by offering wide-area aerial coverage; however, manual examination of drone imagery is time-consuming and susceptible to human error. In this work, a lightweight human detection network based on YOLOv11, referred to as YOLOv11-Star, is proposed for real-time maritime SAR applications. The model substitutes Starblock modules as the backbone feature extractors in the original backbone network, resulting in a significant reduction in computational complexity while maintaining the spatial representations necessary for human detection.
A comprehensive preprocessing pipeline that included environmental augmentation, adaptive tiling, and resplitting was used in image situations. YOLOv11-Star was compared with the baseline YOLOv11, YOLOv11-Starnet and other YOLO models on detection and computational performance. The experimental results demonstrate that the proposed model achieved a 54% parameter reduction while maintaining a compact size of 2.53 MB and generating lightweight detection results of only 0.15-0.16 KB per frame, facilitating UAV-to-ground transmission. Despite the slight decrease in recall, the model achieves 0.806 in precision and 0.714 mAP@0.50, demonstrating competitive performance in detecting visible swimmers while maintaining suitability for real-time deployment.
All things considered, YOLOv11-Star offers a good balance between speed and accuracy, providing a strong foundation for its incorporation into SAR and UAV-based maritime surveillance operations.
Abstract Format
html
Abstract (Filipino)
Ang paghahanap ng mga taong nasa panganib ay nagiging lubhang problematiko dahil sa laki at hindi mahuhulaang katangian ng kapaligirang pandagat, lalo na kapag lumalala ang bisibilidad at nagiging napakaliit ng mga bagay kumpara sa ibabaw ng tubig. Ginagamit ang mga Unmanned Aerial Vehicles (UAVs) upang mapabilis ang mga operasyon ng search and rescue (SAR) sa pamamagitan ng pagbibigay ng malawak na saklaw mula sa himpapawid; gayunpaman, ang manu-manong pagsusuri ng mga larawang kuha ng drone ay matagal at madaling magkaroon ng pagkakamali ng tao. Sa pag-aaral na ito, isang magaan na human detection network na nakabatay sa YOLOv11, na tinatawag na YOLOv11-Star, ang iminungkahi para sa real-time na maritime SAR applications. Pinalitan ng modelo ang mga Starblock module bilang backbone feature extractors sa orihinal na backbone network, na nagresulta sa malaking pagbawas sa computational complexity habang pinananatili ang mga spatial representation na kinakailangan para sa human detection.
Gumamit din ng isang komprehensibong preprocessing pipeline na kinabibilangan ng environmental augmentation, adaptive tiling, at resplitting sa mga sitwasyon ng imahe. Inihambing ang YOLOv11-Star sa baseline YOLOv11, YOLOv11-Starnet, at iba pang YOLO models batay sa detection at computational performance. Ipinakita ng mga resulta ng eksperimento na nakamit ng iminungkahing modelo ang 54% na pagbawas sa parameters habang pinananatiling compact ang laki nito sa 2.53 MB, at nakabuo ng magagaan na detection outputs na 0.15–0.16 KB bawat frame, na nakatutulong sa UAV-to-ground transmission. Sa kabila ng bahagyang pagbaba sa recall, nakapagtala pa rin ang modelo ng 0.806 precision at 0.714 mAP@0.50, na nagpapakitang nakikipagsabayan ito sa pag-detect ng mga nakikitang swimmer habang nananatiling angkop para sa real-time deployment.
Sa kabuuan, ang YOLOv11-Star ay nag-aalok ng balanseng kombinasyon ng bilis at katumpakan, at nagbibigay ng matibay na pundasyon para sa integrasyon nito sa mga operasyon ng SAR at UAV-based maritime surveillance.
Abstract Format
html
Language
English
Format
Electronic
Keywords
Image processing; Neural networks (Computer science)
Recommended Citation
Satore, J. C. (2025). Enhancing YOLOv11 for human detection in maritime environments: Optimization and performance evaluation. Retrieved from https://animorepository.dlsu.edu.ph/etdm_physics/28
Upload Full Text
wf_yes
Embargo Period
12-12-2026
