|
Identifier |
000434012 |
Title |
Machine learning techniques for the detection of illegal human activity in audio recordings from protected areas |
Alternative Title |
Χρήση τεχνικών μηχανικής μάθησης για τον εντοπισμό ανθρωπογενούς δραστηριότητας από ηχογραφήσεις σε προστατευόμενες περιοχές |
Author
|
Ψαρουλάκης, Κωνσταντίνος Σ.
|
Thesis advisor
|
Τσακαλίδης, Παναγιώτης
|
Reviewer
|
Στεφανάκης, Νικόλαος
Στυλιανού, Γιάννης
|
Abstract |
Human activity is considered today as the primary reason for habitat loss for a large number of Earth's
plant and animal species. This activity results to the permanent loss of species and to the weakening of
the ecosystems that are of significant importance for the overall health of the planet and as a
consequence, to the quality of the human life. One key measure to protect habitats is the establishment
of protected areas where human activity is restricted. In these areas, systems employing multiple
cameras and microphones may offer a significant assistance in monitoring the health of the ecosystem
but also as the means to prevent human intervention that is harmful to the environment.
This Thesis concerns the application of signal processing and machine learning techniques to audio
recordings acquired in protected areas in Greece, with the aim to automatically detect sound events
that are indicative of illegal human activity such as illegal logging, grazing, hunting, etc. To collect and
annotate the data that is required for training such a scheme, we illustrate the usefulness of a Voice
Activity Detector (VAD) that is activated on the presence of harmonic structure in the audio content. The
VAD is used in order to automatically segment hundreds of hours of audio recording into thousands of
short duration audio clips that potentially carry the underlying pattern of interest.
Continuing, we perform numerous experiments with the goal to find the optimal approach for (i) a
binary classification problem that focuses on the case of chainsaw sound and (ii) a six class problem that
includes additional patterns relating to illegal human activity. Experimental results illustrate the
superiority of Deep Neural Networks (DNN) against other well-known conventional classifiers and
furthermore, highlight choices that are advantageous for the intended task in terms of the DNN
architecture and the type of acoustic features.
|
Language |
English |
Subject |
Chainsaw |
|
Classification |
|
Neural network |
|
Voice |
|
Ήχος |
|
Αλυσοπρίονο |
|
Ανίχνευση |
|
Εντοπισμός |
|
Ηχογράφηση |
|
Μηχανική μάθηση |
|
Νευρωνικά δίκτυα |
|
Παράνομη ανθρωπογενής δραστηριότητα |
|
Ταξινόμηση |
Issue date |
2020-11-27 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Views |
585 |