COOPMamba: Efficient Vehicle-to-Vehicle Cooperative Perception Based on 3-D Point Clouds

Vehicle-to-vehicle (V2V) cooperative perception promises enhanced environmental awareness but still suffers from two critical gaps: 1) existing intermediate-fusion models typically rely on single-branch feature aggregation, which struggles to simultaneously capture localized discriminative cues and long-range contextual structures and 2) attention-based fusion modules, like Transformers, incur high computational and communication overhead, limiting real-time deployment. To address these limitations, we propose COOPMamba in this article, a lightweight yet expressive collaborative-perception framework that introduces a dual-branch decomposition of salient and global information and a CMamba cross-branch state-space interaction block. The proposed CMamba block performs direction-aware, long-range feature propagation, enabling more robust multivehicle feature alignment and mitigating missed detections caused by incomplete observations. Extensive experiments on V2XSet and OPV2V demonstrate the effectiveness of our design: COOPMamba achieves 89.1% AP@0.5 and 77.1% AP@0.7 on V2XSet, outperforming state-of-the-art (SOTA) approaches by 2.5%-6.4% while maintaining the lowest MACs and inference latency among existing fusion methods. The results confirm that our cross-branch state-space modeling substantially improves collaborative 3-D object detection under both ideal and noisy real-world conditions. The source code is available at https://github.com/npunancy/coopmamba