Latency-Adaptive Feature Fusion Weight Allocation Under Bandwidth Constraints for V2X Cooperative 3D Object Detection

Abstract

Vehicle-to-Everything (V2X) cooperative perception enhances autonomous driving safety by fusing sensor data from multiple agents, including vehicles and roadside units. Communication latency and limited bandwidth remain two critical challenges that jointly degrade fusion accuracy in real-world deployments. Existing research has addressed delay compensation and bandwidth-efficient transmission as separate problems, leaving the coupled impact of these two constraints on fusion performance insufficiently explored. This paper investigates a latency-adaptive feature fusion weight allocation strategy under bandwidth-constrained V2X communication conditions. A temporal decay function is formulated to quantify the degradation of information reliability caused by varying communication delays, and a spatial relevance scoring mechanism is designed to prioritize high-value features when available bandwidth is limited. The proposed weight allocation approach integrates temporal and spatial dimensions to dynamically adjust the fusion contributions of each cooperative agent. Experiments are conducted on three public cooperative perception datasets—DAIR-V2X, V2X-Sim, and V2V4Real—under simulated latency ranging from 0 to 500 ms and bandwidth constraints from 0.04 to 2.0 Mbps. Results demonstrate that the proposed approach achieves consistent improvements of 2.1–4.7% in Average Precision over delay-unaware baselines, with marginal computational overhead suitable for real-time deployment.