Abstract:
The applications of drones are heavily dependent on visual information and
often neglect acoustic information. However, a drone audition is essential, es-
pecially when an object obstructs the view of the target, camera frame restricts
the view, or in an ill-illuminated environment. The main problem of drone au-
dition is the drone's rotor and motor that is very noisy, resulting in a very low
signal to noise ratio (SNR). The low SNR is problematic for solving the sound
source localisation problem for applications such as automatic camera tracking
or search and rescue missions. To improve the drone audition, researchers inves-
tigated microphone array designs, noise reduction techniques, and localisation
algorithms. Microphone array design in particular had issues where researchers
were increasing the overall drone size and weight to improve the SNR. How-
ever, this also increases the rotor noise and the di culty to transport or control
the drone. A recently development of a long rectangular cuboid microphone
array suggests a more viable array design than other conventional shapes. It is
lightweight and can improved the SNR of the observed signal without signi -
cantly a ecting the size and weight of the drone. The main issue of this array
is that it has not been thoroughly researched. Therefore, this thesis investi-
gates the use of a rectangular cuboid shaped microphone array for estimating
the direction of arrival (DOA) of a sound source. To improve the accuracy of
localisation algorithm, a deep convolutional de-nosing autoencoder with Wiener
lter were used. The performance of the array and noise reduction algorithm
were evaluated from the accuracy of the a baseline localisation method.