Distributed 3D Collaborative Perception Mechanism Based on Object-Level Variable Bitrate Protocol in Vehicular Networks

Abstract

Collaborative perception addresses the physical limitations of single-agent sensing in autonomous driving by utilizing Vehicle-to-Everything communication architectures to exchange complementary sensor observations. However, transmitting high-dimensional feature maps or raw point clouds routinely precipitates network congestion, particularly in dense traffic scenarios subject to bandwidth constraints. To resolve the trade-off between perception accuracy and communication overhead, this research introduces a distributed 3D collaborative perception mechanism governed by an Object-Level Variable Bitrate protocol. By evaluating the spatial-temporal heterogeneity of perceptual information, the framework quantifies the task-specific confidence of individual spatial features to execute foreground-background differentiation. An object-supervised confidence head first estimates the task relevance of each BEV cell, after which a multi-level variable-bit-depth codec assigns different quantization precisions to object and contextual regions. Detection loss, confidence supervision, entropy-estimated communication rate, and quantization distortion are jointly optimized, while the actual communication cost is measured from the encoded bitstream. The variably compressed features are subsequently processed using conventional spatiotemporal alignment and attention-based multi-agent aggregation to evaluate the proposed codec under asynchronous cooperative-perception conditions. Experimental evaluations on OPV2V and V2X-Sim demonstrate that the proposed method achieves a favorable detection–communication trade-off under restrictive bandwidth conditions. At an aggregate communication rate of 0.50 MB/s, the proposed method achieves 78.91% AP@0.5 and 64.25% AP@0.7, outperforming the strongest communication-efficient baseline by 5.05 and 6.51 percentage points, respectively. These results provide a practical foundation for bandwidth-aware collaborative perception in dense vehicular networks.