E-Locus - Institutional Repository of the University of Crete

Home Search

Results - Details

Search command : Author="Μουχτάρης" And Author="Αθανάσιος"

Current Record: 8 of 29

[Add to Basket]

Identifier

000408512

Title

Automatic organization of user generated content based on audio correlations

Alternative Title

Αυτόματη οργάνωση περιεχομένου προερχόμενου από πληθοπωρισμό βάσει συσχετίσεων στο ηχητικό υλικό

Author

Χωνιανάκης, Σταύρος Π.

Thesis advisor

Μουχτάρης, Αθανάσιος

Reviewer

Τσακαλίδης, Παναγιώτης
Αργυρός, Αντώνιος

Abstract

With the proliferation of smart-phones and portable electronic devices, more and more of us become engaged in the process of capturing and sharing audiovisual content from public events that we attend. Such User Generated Content (UGC) can be very valuable to the broadcasters and producers associated to the professional coverage of such events, as it may enrich the footage or provide coverage for parts of the event that have not been captured by the professional equipment. Yet, it is not trivial to organize this content in a way that it can be usable for the said purpose. For example, as user generated content lacks metadata which is informative about the exact location and time of recording, it would require enormous time and effort from the professional editor to manually search for videos referring to a particular segment of the event, or to group and temporally align multiple videos overlapping in space and time. Fortunately, as several works have demonstrated in the past, it is possible to automatically organize such content by exploiting the correlations in the audio streams available in the UGC. Rather than working with the raw audio data, such correlations are much more efficiently revealed based on fingerprints extracted from each user generated audio recording. Audio fingerprinting finds use as the means for providing a compact and concrete content based signature, by retaining the maximum acoustically relevant information, showing significant robustness to variations with respect to the audio format, the induced noise and distortion in each audio recording. In this thesis, we use audio fingerprint cross-correlation as the means to detect and synchronize temporally overlapping user generated audio recordings of the same event and we evaluate our tools based on two datasets that we have acquired ourselves; a musical concert and a football match. We perform an extensive comparison based on different state of the art fingerprinting techniques and we propose a novel fingerprinting algorithm with significantly better organization performance for the case of the athletic event. We propose a generalizable scheme for fingerprint cross-correlation and we present an approach inspired by graph theory for clustering recordings from the same temporal segment of the public event into the same group.

Language

English

Subject

Audio clustering and matching

Audio fingerprinting

Audio sychronization

Αποτυπώματα ήχου

Ομαδοποίηση αρχείων ήχου

Συγχρονισμός αρχείων ήχου

Issue date

2017-03-17

Collection