Intelligent Execution of Computer Vision Tasks in Delay-Constrained UAV-aided Networks

Autonomous unmanned aerial vehicles (UAVs) are crucial in critical target tracking and disaster management ser- vices. However, challenges arise due to limited channel capacity causing large transmission delays and constraints imposed by UAV batteries when running computationally intensive object detection and tracking algorithms. To address this, we propose in- telligent offloading computer vision tasks to a high-computational edge server at millimeter wave (mmWave) frequency. Transmis- sion at mmWave needs large transmission power and is suscep- tible to blockages that necessitate considering the link quality in the offloading policy. Additionally, the timely processing of frames containing objects of interest is essential for context-based UAV operations to achieve a low frame drop rate. In this work, we present a delayed-reward reinforcement learning framework to determine the offloading policy of computer vision tasks in a delay-constrained environment. This approach considers the importance of the content within frames, which is unknown to the UAV. The objective is to jointly reduce UAV energy consumption and frame drop rates, leveraging statistical information of both the channel and the frame semantics. Through extensive sim- ulations, we demonstrate by considering statistical information of the communication channel and frame semantics, we achieve approximately 45% energy savings compared to the UAV’s energy consumption when processing all frames locally and maintaining the drop rate of delay-constrained frames below 5%.