Abstract:
Natural and man-made disasters are unexpected events that concern nations
worldwide. Every year, disasters result in infrastructural damages, monetary
costs, population displacement, injuries, and deaths. Unfortunately, climate
change strengthens the destructive power of disasters. The first 72 hours
are crucial for rescuing survivors after the occurrence of the disaster. Lack
of appropriate and rapid assistance to disaster victims due to ineffective
identification and assessment of disaster struck areas makes it very difficult
to plan and execute rescue missions. Unmanned Aerial Vehicles (UAVs) with
embedded Graphical Processing Units (GPUs) and high resolution cameras
provide ease in remotely sensing the disaster area and autonomy in decision
making in disaster recovery and management. However, on-board embedded
GPUs have some limitations in terms of computational power, storage, power
consumption, and size. In this research, we improved cutting-edge image
classification models for effective on-board integrated GPU implementation.
The outcomes of the experiments show that the optimized compressed model
is up to 92 times faster than the native model, which has a throughput
improvement of almost 99 percent. Additionally, the proposed architecture
allows for a model size reduction of roughly 84 percent without sacrificing
accuracy.
