Distributed inference in IoT-based aerial network of UAVs

Abstract

Interest in unmanned aerial vehicle (UAV) has increased drastically due to its ability as a mobile internet of things (IoT) device and capability of accessing remote regions. Due to its robust capabilities with the additional compute capabilities of the onboard IoT device, its tasks range widely from monitoring smart farms to surveillance and rescue related missions. The rise in intelligent tasks has led to an increase in required computing resources as more of these tasks utilize deep neural networks (DNNs) for inference. However, the computing resources of these IoT devices are fixed and limited onboard an UAV and require additional assistance. Some works have suggested UAV–edge collaboration, yet face restrictions for remote deployment. Other works have suggested UAV-to-UAV collaboration, but have not considered simultaneously occurring requests for multiple DNNs and heterogeneous IoT hardware onboard UAVs. Motivated by the need for performing critical tasks in remote areas with heterogeneity in the requested DNNs, we propose a deployment which considers such information during planning. We test our proposed deployment with several others methods to compare e2e latency performance and energy efficiency. Proposed deployment demonstrates lower e2e latency and greater energy efficiency than all other comparison methods.