Abstract:
"Defining virtual fences for crossing detection plays a pivotal role in diverse applications, ranging
from security surveillance to location-based services. There must exist a well-defined coordinate
system and a reliable GPS or other location based services to implement any boundary crossing
detection system. While geometry-based algorithms are traditionally used to detect if a point is
inside or outside a geofence, applying this process for large numbers of points and for complex
geofences is resource intensive. Further, when a real time operation is needed, this may lead to
CPU wastage due to the continuous processing of large numbers of individual points. Some
assets such as space objects or micro level organisms cannot be easily tracked with a solution
such as GPS, as these do not exist in a defined coordinate grid nor implanting a tracker device in
such objects is feasible.
This research focuses on a novel approach to introduce image analysis for boundary crossing
detection. This approach primarily addresses the limitation of complex geofences and the
exhaustive CPU resource usage when monitoring many points. The author further introduces a
free hand sketched geofence boundary annotation system to the environment of interest opening
up the use of geofence in non-coordinate systems. The proposed system, ‘Geogle’, uses a deep
learning model trained on a dataset generated by the author.
Final test results prove the system has an accuracy of 92% with the generated dataset and the
phase 2 model. The system is able to monitor complex geofences as well as allow a user to
define free-hand sketched geofences using the UI. A separate canvas drawing module is
integrated for this purpose."
