DRL-Based Joint Task Processing and Resource Allocation for Remote Sensing in Vehicle-Device-Edge Collaborative Networks

Abstract

Remote sensing is a key enabling technology for intelligent vehicle and transportation systems, as it acquires environmental and traffic data beyond the vehicle’s visual range. However, in practical scenarios, the lack of flexible collaboration between sensing nodes and computing facilities leads to low efficiency in the link from remote data sensing to inference, making it difficult to support the effective processing of remote data and meet the requirements of real-time tasks. To address this issue, this paper proposes a vehicle-sensing-edge collaborative model for AI task scheduling and resource management, which integrates multiple vehicles, various sensing devices, and multi-RSU edge servers. This model performs integrated management of remotely collected data and subsequent related AI task scheduling and computing processes, enabling the flexible execution of AI inference tasks on task vehicles, service vehicles, sensing devices, and RSU edge servers. By collaboratively optimizing data collection paths, communication links, and computing resource allocation, the system achieves the joint optimization of task inference accuracy and overall operational cost while balancing load equilibrium and low latency. We design a hybrid algorithm based on deep reinforcement learning (DRL), which decomposes the original problem into sub-problems and combines the Softmax Deep Double Deterministic Policy Gradient (SD3) algorithm with convex optimization calculations. The computational complexity and convergence performance of the proposed algorithm are analyzed in detail. Theoretical analysis demonstrates that the algorithm’s inherent advantages of low complexity and fast, stable convergence guarantee the effectiveness of problem-solving. Meanwhile, comparative experiments with 5 comparison schemes under 6 typical scenarios are conducted, which fully verify the superiority of the proposed scheme.