logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-06072020-160831


Tipo di tesi
Tesi di laurea magistrale
Autore
SENORE, ELISABETTA
URN
etd-06072020-160831
Titolo
Generalization Capabilities of Deep Learning Architectures for Vehicle Re-Identification
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Relatori
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Prof.ssa Vaglini, Gigliola
relatore Ing. Celentano, Giovanni
Parole chiave
  • cnn
  • convolutional neural network
  • deep learning
  • generalization capability
  • re-identification
  • resnet
  • resnet-50
  • vehicle
Data inizio appello
22/06/2020
Consultabilità
Completa
Riassunto
Vehicle re-identification is a computer vision problem that consists in recognizing the same vehicle in the traffic under different perspectives. Vehicle re-identification finds broad applications in urban scenarios, for example in the estimation of the travel time, in the management and modeling of the traffic system, or in seeking and tracking vehicles of interest for reasons of security. The license plate is not always recorded, therefore most of the vehicle re-identification models rely on the appearance of the vehicle. Nevertheless, if a model is trained for a specific set of cameras it is hard to make it work properly on cameras with different viewing conditions, i.e. to make the model general and not strongly dependent from the context. In this study, different models at the state-of-the-art were analyzed, and two of them were implemented for testing their generalization ability - Vehicle ReID Baseline and Deep Meta Metric Learning (DMML). Both Vehicle ReID Baseline and DMML employ a particular type of convolutional neural network, the ResNet-50, that is the backbone of the model. The neural network was trained on the VeRi dataset. The generalization ability of the models was tested on VehicleID dataset.
File