Energy-Efficient Collaborative DNN Inference in UAV Swarm.

Unmanned Aerial Vehicles (UAVs) have attracted great attention due to its high mobility and flexible deployment, which makes many people develop UAVs for different application scenarios. The unique performance of UAVs encourages the emergence of more critical and complex tasks in uncertain and potential harsh environments, many of which were not even envisaged decades ago, including military border surveillance and oil/gas offshore exploration. The large amount of data generated by these applications needs to be processed and analyzed by deep neural networks (DNNs). However, due to the resource constraints of UAVs, there are many challenges in how to deal with deep networks and complex models on UAVs. In this paper, we propose a fine-grained DNN inference partition method, aiming at allocating inference requests to the resource-constrained UAV swarm, classifying the captured images, and finding the minimum inference latency. We formulate the problem as an optimization problem that minimizes the inference latency of UAV swarm from capturing the image to outputting the result. The optimization problem is an NP-hard problem. Therefore, we introduce an online heuristic solution, namely EECIA, to find the computing task allocation strategy that gives the best latency among the available UAVs. The simulation results show that compared with the benchmarks, our algorithm can get relatively good performances under different configures and the running time meets the real-time requirements.