<div>
<div>
<div>
<p>This paper introduced a near real-time acoustic unmanned aerial vehicle detection system
with multiple listening nodes using machine learning models. An audio dataset was
collected in person by recording the sound of an unmanned aerial vehicle flying around as
well as the sound of background noises. After the data collection phase, support vector
machines and convolutional neural networks were built with two features, Mel-frequency
cepstral coefficients and short-time Fourier transform. Considering the near real-time
environment, the features were calculated after cutting the audio stream into chunks of
two, one or half seconds. There are four combinations of features and models as well as
three versions per combination based on the chunk size, returning twelve models in total.
To train support vector machines, the exhaustive search method was used to find the best
parameter while convolutional neural networks were built by selecting the parameters
manually. Four node configurations were devised to find the best way to place six
listening nodes. Twelve models were run for each configuration, generating color maps to
show the paths of the unmanned aerial vehicle flying along the nodes. The model of
short-time Fourier transform and support vector machines showed the path most clearly
with the least false negatives with 2-second chunk size. Among the four configurations,
the configuration for experiment 3 showed the best results in terms of the distance of
detection results on the color maps. Web-based monitoring dashboards were provided to
enable users to monitor detection results. </p>
</div>
</div>
</div>
Identifer | oai:union.ndltd.org:purdue.edu/oai:figshare.com:article/7975991 |
Date | 10 June 2019 |
Creators | Bowon Yang (6574892) |
Source Sets | Purdue University |
Detected Language | English |
Type | Text, Thesis |
Rights | CC BY 4.0 |
Relation | https://figshare.com/articles/UAV_DETECTION_SYSTEM_WITH_MULTIPLE_ACOUSTIC_NODES_USING_MACHINE_LEARNING_MODELS/7975991 |
Page generated in 0.0022 seconds