Semantic-driven Performance Optimization for UAV Based Object Detection

Unmanned aerial vehicles (UAVs) often need to capture and transmit videos to accomplish intelligent tasks. Minimizing the channel pressure from video transmission while ensuring task quality is a crucial challenge. Semantic communication is envisioned as a new paradigm for future intelligent task-oriented applications in 6G networks. However, the existing semantic communication frameworks require devices independently performing numerous computations to extract semantic features, which is unbearable for energy-limited UAVs. In this paper, we propose an efficient semantic communication framework for the object detection. Our proposed framework is able to address this issue by compressing transmitted data and offloading complex computational tasks to an adjoining edge server. This offloading strategy significantly reduces the computational loads for the UAVs. To measure the performance of the proposed framework, we propose a new semantic metric named quantity of detected objects (QoDO), which jointly considers the impact of video acquisition and transmission processes on the number of targets detected. Subsequently, we propose a semantic driven optimization scheme for UAV trajectory design and power allocation. To solve this problem, a hierarchical deep reinforcement learning algorithm based on two Deep Deterministic Policy Gradient (DDPG) networks (H2DDPG) is proposed. Simulation results show that the proposed framework and scheme achieve a steady improvement in the total QoDO compared to the existing schemes that without considering semantics. The proposed semantic driven optimization method is practical in solving performance bottlenecks for intelligent task-oriented communication systems.