Distributed Intelligence Enabled Multi-Vehicle Collaborative Perception: Latency-Accuracy-Stability Trade-Offs

Abstract

Multi-vehicle collaborative perception (MvCP) is a promising paradigm to enhance the capability of autonomous driving (AD). However, meeting the low latency and high accuracy requirements in a distributed, dynamic, and heterogeneous vehicular network is challenging. In this paper, we formulate a distributed framework enabling MvCP, where a <underline>d</underline>elayed-<underline>f</underline>usion <underline>par</underline>allel <underline>infer</underline>ence with model <underline>part</underline>ition (DF-PartInfer) scheme is proposed based on the theoretical analysis for <underline>p</underline>arallel <underline>s</underline>eparability (PS). Moreover, we design a metric to characterize the <underline>l</underline>atency, <underline>a</underline>ccuracy, and topology <underline>s</underline>tability <underline>t</underline>rade-offs (LAST). To maximize LAST, we propose a <underline>j</underline>oint <underline>c</underline>ollaborative <underline>v</underline>ehicle (co-vehicle) selection, <underline>m</underline>odel <underline>p</underline>artitioning and <underline>r</underline>esource allocation (JCV-MPR) algorithm, which is designed as a two-layer structure to decouple the total problem into integer and continuous sub-problems. Specifically, in the outer layer, we design the graph neural network (<underline>G</underline>NN)-based multi-agent proximal policy optimization (<underline>M</underline>APPO) co-vehicle <underline>s</underline>election and model <underline>p</underline>artitioning (GMSP) algorithm to determine the integer variables of co-vehicle selection and model partitioning. With the integer variables fixed, we design the <underline>j</underline>oint <underline>c</underline>ommunication and <underline>c</underline>omputation resource <underline>a</underline>llocation (JCCA) algorithm in the inner layer for continuous resource allocation. Simulation results show that the proposed DF-PartInfer scheme and JCV-MPR algorithm outperform the baselines, reducing latency by up to 58.96%.