Efficient Vehicle Recognition and Tracking for UAV-Enabled Intelligent Transport Systems: A Multi-Agent Reinforcement Learning Method

【Author】 Ouyang, Wenjiang; Mu, Junsheng; Jing, Xiaojun; Wang, Yi

【Source】IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

【影响因子】9.551

【Abstract】Vehicle recognition constitutes a foundational technology within intelligent transport systems (ITS), enabling real-time recognition, classification, and tracking of vehicles. With the characteristics of low construction cost, flexible deployment and strong environment adaptability, unmanned aerial vehicle (UAV) is increasingly leveraged for vehicle target recognition, acts as the air part of future intelligent transport systems (ITS) for traffic management, accident handling and vehicle order management, and provides a more efficient, safe and sustainable transport mobility solutions in future ITS. Promoted by the massive number of intelligent vehicles and growing demands of connected vehicles in ITS, continuous and high-fidelity spatio-temporal monitoring of vehicle movement is expected in future ITS, raising the pursuit of higher vehicle recognition performance. As a typical distributed training framework, federated learning (FL) is a desired paradigm to improve sensing performance with the communication of sensing parameters for UAV-enabled ITS. Due to the heterogeneity of sensing data in the cooperative UAV-enabled ITS, the non-independent identically distribution (Non-IID) issue is inevitable. The existing data augmentation works aimed at Non-IID issue in FL utilize single-agent reinforcement learning (SARL), where the local model parameters are input into a central network, resulting in the model privacy leakage problem. To deal with the above issue, a multi-agent reinforcement learning (MARL) algorithm is applied to optimize the training accuracy and data augmentation efficiency for UAV in ITS. Moreover, a decentralized blockchain-based FL (BFL) framework is proposed to avoid the single-point failure in UAV-enabled ITS. The experiments are conducted on the generated vehicle dataset (VRID) and the simulation results indicate that our proposed algorithm exhibits a superior performance than the benchmark algorithms, especially in terms of higher vehicle target recognition accuracy and lower communication overhead, which provides a significant technology support for vehicle identification and tracking in future ITS.

【Keywords】Training; Autonomous aerial vehicles; Sensors; Data models; Target tracking; Privacy; Reinforcement learning; Monitoring; Data privacy; Accuracy; Intelligent transport systems; federated learning; multi-agent reinforcement learning; unmanned aerial vehicle; vehicle recognition and tracking

【发表时间】2025 2025 SEP 1

【收录时间】2025-09-15

【文献类型】

【主题类别】

--