Gunshot Detection in Wildlife using Deep Learning

Master thesis (2021)

Authors

D. Danaei Electrical Engineering, Mathematics and Computer Science

Contributors

Marco Zuniga Embedded Systems - EEMCS (supervisor 1)
J. Yang Web Information Systems - EEMCS (supervisor 2)
Arjan Smit (coach)
Babak Ghafari (coach)
Faculty
Electrical Engineering, Mathematics and Computer Science
Copyright: © 2021 Deniz Danaei
More Info
expand_more

Published Date

23-12-2021

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Electrical Engineering, Mathematics and Computer Science

Abstract


The fight against the illegal hunting of African wildlife is a never-ending process. In order to preserve animal habitats and save them from extinction, many national parks utilize surveilling solutions to prevent, detect and locate intruders. One strategy to detect and locate the illegal hunters or so-called extit{poachers} is to detect and locate the gunshot sounds using an acoustic surveillance system consisting of embedded devices scattered within the park. The embedded devices–so-called end-nodes surveil the environment continuously, processing the sound events gathered and converted to audio by the acoustic sensors. Then, using a deep learning algorithm, any sound event classified as a gunshot is reported to the authorities.

This research study proposes a deep learning model for gunshot sound recognition in African wildlife. It also investigates a potential correlation between gunshot sound recognition accuracy, signal-to-noise ratio (SNR), and shooter distance.

To this end, gunshot and ambient sounds such as Savanna wildlife, rain, and thunder were collected and synthesized to simulate different scenarios. Various experiments were conducted using this data to investigate the influence of different parameters on gunshot recognition accuracy.

Our analysis revealed the negative effect of the weather conditions, such as rain and thunderstorms, on the model accuracy.
The obtained results also showed a positive correlation between the gunshot recognition accuracy and SNR. Since SNR is negatively correlated to the shooter distance when both noise and signal levels are constant, we have proved that the gunshot recognition accuracy also negatively correlates with the distance.

Finally, a single CNN (convolutional neural network) model is proposed for gunshot sound recognition in African wildlife. The model performs acceptably in three different weather conditions. The gunshot recognition accuracy for five shooting ranges is also provided based on the uncovered correlation.