HAP-UAV-assisted hierarchical aerial computing framework for video offloading: a deep reinforcement learning approach

Unmanned aerial vehicles (UAVs) and high altitude platforms (HAPs) can work together to provide real-time video analytics for ground camera equipments (CEs) in disaster rescue. Resource-constrained UAV can offload video analytics tasks to HAP to optimize computation delay, but long-distance wireless video transmission from UAV to HAP with limited bandwidth leads to unacceptable transmission delay, making it necessary to transcode video into a lower bitrate before transmission. However, bitrate compression degrades video quality and sacrifices image detail, which reduces video analytics accuracy. To improve the tradeoff between delay, including transmission and computation delay, and video quality for video offloading, this paper develops a hierarchical aerial computing framework consisting of multiple UAVs and an HAP. Firstly, we employ an adaptive UAV-HAP cooperative computing and video transcoding scheme. Each UAV collects video from CEs within its coverage and selects the optimal offloading ratio, one portion is transcoded into an appropriate bitrate, then offloaded to HAP, and the other is computed locally. Secondly, we build a quality of experience (QoE) model as the optimization objective and jointly optimize offloading ratio, transcoding ratio and HAP computation resource allocation. This nonconvex problem is modeled as a Markov decision process (MDP). To achieve continuous action control, we propose a deep deterministic policy gradient (DDPG) algorithm to solve this optimization problem. Simulation results validate the superiority of our proposed framework and algorithm in balancing delay and video quality for video offloading in aerial computing scenarios.