Object Recognition Offloading in Augmented Reality Assisted UAV-UGV Systems

Abstract

A multi-UAV-UGV system combines the advantages of both UAVs and UGVs, hence it can be used for challenging missions. When incorporated with Augmented Reality (AR), such a system can better involve humans in the loop to either provide feedback to robots' plans or make informed decisions. One common problem in such a system is object recognition. To conserve energy on the UAVs, offloading part of the computation in object recognition is considered. In this paper, we propose and implement two offloading techniques. We compare them against two baselines: zero offloading (i.e., all local computation) and full offloading (i.e., all offloaded to UGV) by implementing these strategies on the three most popular onboard computers on UAVs: Raspberry Pi 4, Jetson Nano, and Jetson Xavier NX. We use a laptop to represent the onboard computer on a UGV. Our experimental results validate the feasibility and benefit of object recognition task offloading in a multi-UAV-UGV setting and also highlight the need for a more effective offloading strategy.