Toward collaborative perception for CAVs: A joint optimization of latency, multimodel fusion, and resource allocation

Abstract

In complex traffic environment, the limited sensing range of Connected and Autonomous Vehicles (CAVs) leads to missing Regions of Interest (ROI) when occlusions occur. Collaborative perception can extend the sensing range and recover ROI through information sharing, but it also introduces additional computation and increases processing latency. In this paper, we propose a joint latency-resource optimization architecture to achieve accurate and real-time perception under occlusions. First, we model an edge-vehicle collaborative network, and formulate a dual optimization problem for resource allocation and minimizing perception data collection latency. Second, with resource constraints, this problem is transformed into an ROI fusion quality optimization problem, which is solved by the Improved Discrete Salp Swarm (IDSS) algorithm. Finally, based on the fusion quality optimization results, the Hybrid Action Space-based Deep Deterministic Policy Gradient (HDDPG) algorithm is introduced to dynamically allocate available computation resources, thereby achieving minimized perception fusion latency. Simulation results demonstrate the efficiency of the proposed joint optimization architecture, which could reduce the processing latency by 0.1 s compared to comparison algorithms, with higher accuracy.