Sharing the Edge: System Status Aware Object Recognition Task Offloading

The combination of object recognition capability with Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) benefits applications like remote surveillance, search and rescue, and infrastructural monitoring. Offloading deep learning-based object recognition models from UAVs to UGVs can address energy and computational constraints on UAVs. However, when multiple UAVs are in contact with one single UGV, how to provide timely offloading decisions considering various system status information such as dynamic network conditions and remaining energy levels on UAVs remains under-explored. This paper presents our work in this area. Our online offloading decision engine, running on the UAV's onboard computer, dynamically determines offloading decisions and encoding bitrates considering fluctuating network conditions. In addition, our task scheduler running on the UGV prioritizes tasks according to the system statuses of each UAV. We conducted extensive evaluations of our approach in both lab and field settings. Experiments show that our system reduces the end-to-end system latency by up to 88% compared to a previous study that reduces both offload data size and local computation.