Tipo di tesi
Tesi di laurea magistrale
Titolo
Domain Adaptation Strategies for YOLO-Based Object Detection in UAV-Assisted Maritime Search and Rescue Operations
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Parole chiave
- computer vision
- DAC-SDC 2022
- deep learning
- domain adaptation
- domain shift
- fine-tuning
- maritime search and rescue
- MOBDrone
- MSAR
- object detection
- small object detection
- transfer learning
- UAV
- unmanned aerial vehicles
- YOLOv12
Data inizio appello
26/05/2026
Consultabilità
Non consultabile
Data di rilascio
26/05/2096
Riassunto (Inglese)
This thesis investigates the development and optimization of object detection algorithms for Unmanned Aerial Vehicles (UAVs) in Search and Rescue (SAR) missions, with a deep focus in Maritime ones..
Given the critical time window for survival at sea, the study explores the evolution of Artificial Intelligence from standard Convolutional Neural Networks (CNNs) to the latest "Attention-Centric" YOLOv12 architecture.
Using the DAC-SDC 2022 and MOBDrone datasets, three experimental methodologies — Strict (zero leakage), Sparse Injection, and Fine-Tuning — were implemented to address the fundamental challenges of Domain Shift and Small Object Detection (SOD).
Performance evaluation and model comparison were primarily driven by the mAP50 metric, which served as the principal benchmark for assessing algorithm efficacy.
Furthermore, the research prioritizes Recall over Precision, reflecting the operational imperative of minimizing False Negatives in life-saving scenarios.
Results demonstrate that "Nano" models (particularly YOLOv11n and YOLOv12n) provide an optimal balance of accuracy and computational efficiency, enabling scalable swarm deployment for real-time monitoring of vast oceanic areas while overcoming hardware limitations and environmental noise.
