Uncrewed Aircraft Detection from YAMNET Embedding Dataset

Published: 12 July 2024| Version 2 | DOI: 10.17632/5dmcszvym4.2
Contributors:
,

Description

This dataset is the main dataset used by the UA Detection Team here at Embry-Riddle Aeronautical University. It provides the ability to conduct binary drone/no drone classification as well as specific drone sub-type classifcation. Dataset is gathered from field data from drones flown by the Unavail research team. Each clip of audio is fed through YAMNET to generate points in an embedding space. The no_drone directory contains various sounds from ambient noise to cars and jets. drone contains samples from the DJI Matrice M100, the Mavic 3, and the Mavic Mini 2. There are 9108 samples total across all the directories. Each .tfdata segment represents a single second of audio.

Files are not publicly available

You can contact the author to request the files

Steps to reproduce

This is a set of data trained using a neural network. Walk through only the first 2 base directories for binary yes/no classification. These embeddings are generated using YAMNET from Google. More data can easily be created by gathering embedding arrays returned by audio samples passed into the YAMNET neural network. YAMNET can be downloaded here: https://www.kaggle.com/models/google/yamnet Classification for the presence of drones or for subtype detection can be done with dense neural networks. We have found best classification results (96% in testing) when only using 2 dense neural network layers.

Institutions

Embry-Riddle Aeronautical University

Departments

Aerospace Engineering

Categories

Intrusion Detection, Drone (Aircraft)