Abstract:Unmanned aerial vehicles (UAVs) have been widely used in military and civilian fields by carrying out intelligent applications with deep learning technologies, such as battlefield search and rescue; intelligence, surveillance, and reconnaissance (ISR); and smart city monitoring. Due to deeper and more complex models of deep neural networks (DNNs) and limited on-board resources in UAVs, most existing works either rely on cloud-based intelligence by transmitting original data to remote powerful cloud servers and running DNNs inference therein, or involve edge intelligence by executing lightweight DNN models on-board. Unfortunately, the former is faced with unacceptable long transmission latency and possible network instability over air-to-ground links, while the latter is restricted to relatively low accuracy via lightweight models. Although a few recent works propose to collaboratively run complex DNNs inference within a swarm of UAVs to achieve both high accuracy and low latency, they seldom consider the strong confrontation environment of the battle-field, the unreliable air-to-air links among UAVs as well as likely hardware/software breakdowns in UAV airborne computers, which could result in the failure of collaborative inference. This article proposes a novel architecture based on an observation, orientation, decision, and action (OODA) loop, called elastic collaborative edge intelligence (eCoEI), which could fully utilize the resources of computation and storage in a UAV swarm, and enable the swarm to collaboratively perform DNN inference tasks in an elastic manner in the face of any lost links. This article also designs a prototype system of the proposed eCoEI and conducts a proof-of-concept evaluation to validate its feasibility as well as effectiveness. Experimental results show that the eCoEI has an outstanding ability to deal with single point of failure and network fluctuations within a UAV swarm for collaborative edge intelligence. Finally, the main technical challenges are pointed out in the eCoEI, and an outlook on future research directions is provided.

Reliable DNN Partitioning for UAV Swarm

Efficient Pipeline Collaborative DNN Inference in Resource-Constrained UAV Swarm

Multi-UAV Maritime Search and Rescue with DNN Inference Acceleration

A hybrid fast inference approach with distributed neural networks for edge computing enabled UAV swarm

Joint Split Offloading and Trajectory Scheduling for UAV-enabled Mobile Edge Computing in IoT Network

DNN Adaptive Partitioning Strategy for Heterogeneous Online Inspection Systems of Substations

A UAV Swarm Sensing Oriented Distributed Computing Cooperation Scheme

Efficient Real-Time Image Recognition Using Collaborative Swarm of UAVs and Convolutional Networks

Cloud-Edge Inference under Communication Constraints: Data Quantization and Early Exit.

DNN Task Assignment in UAV Networks: A Generative AI Enhanced Multi-Agent Reinforcement Learning Approach

Intelli-Eye: An UAV Tracking System with Optimized Machine Learning Tasks Offloading

Real-time Adaptive Partition and Resource Allocation for Multi-user End-cloud Inference Collaboration in Mobile Environment

Elastic Collaborative Edge Intelligence for UAV Swarm: Architecture, Challenges, and Opportunities

An IPSO-assisted DNN Operation Scheme in UAV-mounted MEC Networks.

A Survey on Deep Neural Network Partition over Cloud, Edge and End Devices

A Novel Adaptive Computation Offloading Strategy for Collaborative DNN Inference over Edge Devices.

Dynamic DNN Decomposition for Lossless Synergistic Inference

A DNN Operation Scheme Based on IPSO for UAV-Assisted MEC Networks

DNN Inference Acceleration with Partitioning and Early Exiting in Edge Computing

Deep Neural Network Task Partitioning and Offloading for Mobile Edge Computing

Distributed DNN Inference with Fine-grained Model Partitioning in Mobile Edge Computing Networks