Vision Based Localization of Drones in a GPS Denied Environment
Abstract
In this thesis, we build a robust end-to-end pipeline for the localization of multiple drones in
a GPS-denied environment. This pipeline would help us with cooperative formation control,
autonomous delivery, search and rescue operations etc. To achieve this we integrate a custom
trained YOLO (You Only Look Once) object detection network, for drones, with the ZED2
stereo camera system. With the help of this sensor we obtain a relative vector from the left
camera to that drone. After calibrating it from the left camera to that drone's center of mass,
we then estimate the location of all the drones in the leader drone's frame of reference. We
do this by solving the localization problem with least squares estimation and thus acquire
the location of the follower drone's in the leader drone's frame of reference. We present the
results with the stereo camera system followed by simulations run in AirSim to verify the
precision of our pipeline.
General Audience Abstract
In the recent years, technologies like Deep Learning and Machine Learning have seen many
rapid developments. This has lead to the rise of fields such as autonomous drones and
their application in fields such as bridge inspection, search and rescue operations, disaster
management relief, agriculture, real estate etc. Since GPS is a highly unreliable sensor, we
need an alternate method to be able to localize the drones in various environments in real
time. In this thesis, we integrate a robust drone detection neural network with a camera
which estimates the location. We then use this data to get the relative location of all
the follower drones from the leader drone. We run experiments with the camera and in a
simulator to show the accuracy of our results.
Collections
- Masters Theses [19643]