With the rapid development of edge computing technology, edge-assisted unmanned aerial vehicle (UAV) networks have become popular, helping with fast and cost-effective data collection in mobile crowdsensing (MCS) environments. This paper investigates the online data collection problem for MCS over an edge-assisted UAV network architecture, where UAVs work to collect the data required by tasks at different on-ground point-of-interests (PoIs) in an autonomous and cooperative manner. Different from conventional edge-assisted UAV networks, edge nodes in our paper help distribute, aggregate, share, and update the task state information (TSI, e.g., if a task has been completed, or if it still requires more data), to support efficient and cost-effective data collection, e.g., avoiding repetitive and ineffective task execution. In particular, a UAV can exchange TSI with an edge node when it is within the signal coverage of the edge node, while making decisions on PoI selection and path planning in an online and decentralized manner. To address the issue of edge-assisted UAV data collection, we propose a multi-agent deep reinforcement learning-based algorithm using personalized training with decentralized executing (PTDE) architecture. Different from the traditional centralized training with decentralized executing (CTDE) architecture, our considered architecture adopts the agent-specific state for critic networks instead of the joint observation, thus achieving effective utilization of environmental information. Furthermore, we propose an observation enhancement algorithm based on artificial potential field (APF). Extensive simulation results demonstrate that our proposed algorithm greatly outperforms baseline algorithms in terms of total profit, data collection ratio, geographical fairness, and energy efficiency.
This paper proposes a dynamic unmanned aerial vehicle (UAV) clustering model for multi-target localization in complex 3D environments, where mobility-aware cluster formation is integrated to enhance collaborative localization accuracy. We derive the Cramér–Rao lower bound (CRLB) for localization performance analysis under measurement and motion-induced uncertainties. To solve the NP-hard clustering problem, we develop the MDQPSO-ASA algorithm, which combines multi-swarm discrete quantum-inspired particle swarm optimization with adaptive simulated annealing, incorporating a repair mechanism to satisfy spatial and cardinality constraints. Simulation results demonstrate the algorithm’s superiority in localization accuracy, computational efficiency, and adaptability to varying UAV/target scales compared to baseline methods. The developed algorithm provides an effective solution for resource-constrained collaborative localization tasks in practical scenarios.
1 January 2026 in IEEE Transactions on Cognitive Communications and Networking