PADIS

Pubblicazioni Aperte DIgitali Sapienza > Informatica e sistemistica "Antonio Ruberti" > INGEGNERIA INFORMATICA >

Please use this identifier to cite or link to this item: http://hdl.handle.net/10805/1241

Title: Learning from the Visual Input Statistics
Authors: CARBONE, ANDREA
Issue Date: 17-Feb-2011
Abstract: The importance of a filtering method or a selective criterion to discriminate salient content in images and video sequences in Computer Vision has been widely recognised in recent years. The need for real time visual input processing is becoming evident for applications in home and service robotics, surveillance, driver or disabled people assistance, virtual and augmented reality. Such an ability will help further processing and enhance perceptual organisation of the observed scenes. Nevertheless it needs to rely on adequate representations of the incoming stimuli, that can reduce redundancy and noise, preserving relevant information and facilitating learning.The research work reported in this thesis has been carried out in the long-range perspective of the construction of a biologically inspired model of attention, according to which a robot or an artificial vision system would be able to focus on (task) relevant portions of the perceived scene. Several approaches are presented outlining different experimental scenarios aimed at demonstrating the gain in performance in the visual search when endowing a robotic mobile platform with an attentive spatial visual guidance. Finally, we show how the rich visual information can be modelled when considering as input human recorded scan\--paths on both large databases of eye-tracked fixations on still images and fixations recorded with a custom wearable gaze tracker. The sparse and high dimensional data that is delivered by the ICA basis is a biologically plausible representation for the information flow in the early stages of visual processing. We use the new coded features to statistically model the distribution of fixated areas w.r.t. ignored locations and to instantiate a learning framework for the development of a computational model of visual saliency able to discriminate potentially interesting areas in new scenes.
URI: http://hdl.handle.net/10805/1241
Appears in PhD:INGEGNERIA INFORMATICA

Files in This Item:

File Description SizeFormat
Andrea CARBONE.pdfTesi di Dottorato3.18 MBAdobe PDF

File del Curriculum Vitae:

CurriculumVitae.carbone.currvitae.pdf 61.22 kBAdobe PDF


This item is protected by original copyright

Recommend this item

Items in PADIS are protected by copyright, with all rights reserved, unless otherwise indicated.

 

Valid XHTML 1.0! DSpace Software Copyright © 2002-2010  Duraspace - Feedback Sviluppo e manutenzione a cura del CINECA