logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-07042023-110204


Tipo di tesi
Tesi di laurea magistrale
Autore
CAMPILONGO, FRANCESCO
URN
etd-07042023-110204
Titolo
Design and Development of Artificial Intelligence Techniques for Detecting People at Sea in aerial Images
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Relatori
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Prof. Gennaro, Claudio
relatore Dott. Ciampi, Luca
relatore Dott.ssa Vadicamo, Lucia
Parole chiave
  • fine-tuning
  • mobdrone
  • neural network
  • object detection
  • people in open water
Data inizio appello
21/07/2023
Consultabilità
Completa
Riassunto
Modern camera-equipped Unmanned Aerial Vehicles (UAVs) can play an essential role in accelerating the localization and rescue of people. To this end, Artificial Intelligence (AI) techniques can be leveraged to automatically understand visual data acquired by drones. This has the potential to reduce search times and ultimately save human lives significantly.

This thesis focuses on the field of object detection, a fundamental computer vision task, with a specific emphasis on its application to detecting people in open water environments from drone-view imagery. These kinds of scenarios are remarkably different compared to the generic ones for many reasons: different points of view from which the images are taken, different shapes of the objects seen from above, and changeable backgrounds due to various colors of water and weather conditions.
To address these challenges, we conduct an experimental evaluation using various state-of-the-art deep neural networks over two datasets of images taken from UAVs at different altitudes, the Sea Drones See and the MOBDrone datasets. Specifically, we consider three popular generic object detectors (VarifocalNet, TOOD, and the latest version of YOLO (v8)), stressing their generalization capabilities and measuring their performances in various experimental settings.

Through our experiments, we achieve notable improvements in detection performance. We started from a mean Average Precision (mAP) of 0.378 for the best performer, VarifocalNet pre-trained on a general-context dataset suitable for object detection, to a mAP of 0.647 using the same neural network fine-tuned over the specific drone-view datasets. Furthermore, we evaluate the performance also in terms of efficiency, showing that the latest version of YOLO demonstrates remarkably faster inference times, approximately nine times faster than VarifocalNet, while also consuming less memory.
Our results highlight the significant potential of using AI techniques to improve search and rescue operations in drone-view imagery, paving the way to their applicability even directly onboard the UAVs.
File