Load-Aware Agent Selection via Delay Prediction in V2X Cooperative Perception (to appear)

Abstract

In V2X cooperative perception, vehicles and roadside units share LiDAR features to overcome occlusion and limited sensing range. However, existing frameworks, such as V2X-ViT, do not account for computational delays in processing this data caused by GPU queue congestion at surrounding agents. Sharing delayed features can degrade perception accuracy and waste limited network bandwidth. In this paper, we propose a load-aware agent selection mechanism that predicts feature latency using an M/G/1 queueing model from GPU load information obtained before full feature transmission. The method requests features only from agents with predicted latency below a predefined threshold. Experiments on the V2XSet dataset show that it reduces unnecessary communication traffic while maintaining or improving 3D object detection performance in high-load scenarios.