Confidence-V2X: Confidence-driven sparse communication for efficient V2X cooperative perception

Abstract

Collaborative perception enables agents to enhance their perceptual capabilities by exchanging feature messages with others. However, prior work has typically focused on individual aspects, such as independently investigating improvements in joint perception or reducing communication burdens, often at the expense of achieving strong overall system performance across varying environments. To address the problem of achieving an optimal balance between the object detection ability and communication load of a model, we propose Confidence-V2X, a novel cooperative perception method, that emphasizes the dynamic gating of the feature exchange strategy as well as the optimization of sparse feature refinement and fusion techniques. In Confidence-V2X, we first refine the given raw perceptual features using confidence maps and perform structured packaging to fully prepare for the subsequent process. Next, the outbound interagent communication procedure for compact data exchange is dynamically gated and uniformly scheduled based on a whitelist. Finally, agents update the sparse features along the temporal dimension and adaptively fuse them in the spatial dimension based on confidence information to obtain the final cooperative perception result. Extensive experiments conducted on three datasets demonstrate that Confidence-V2X achieves superior performance to that of the existing methods across multiple metrics while markedly reducing the imposed communication overhead. Our corresponding code will be released on https://github.com/Rwang0208/Confidence-V2X.