An Efficient Cross-Agent Spatial–Temporal Collaboration Framework for Environmental Perception in IoV

By sharing complementary perceptual information, cross-agent collaborative perception fosters a deeper understanding of the environment. Recent studies mostly utilize CNN or Transformer structures to learn feature representation during feature encoding and fusion, which struggle to handle long-range spatial-temporal features under limited computing and communication resources, resulting in suboptimal perception performance. Holistically modeling the dependencies over extensive spatial areas and extended temporal cues is crucial to enhancing feature quality. To this end, we propose CollaMamba, a resource-efficient cross-agent spatial-temporal collaboration framework. We first optimize feature encoding and fusion pipeline based on the spatial state space model, which effectively captures spatial causal dependencies from both single-agent and cross-agent perspectives, yielding compact and comprehensive intermediate features while maintaining linear complexity. Subsequently, to further improve feature encoding quality, we devise a history-aware feature boosting component that extracts contextual cues from extended historical frames, refining ambiguous features while maintaining low overhead. To ensure the reliability of feature fusion under unstable communication conditions, we develop a cross-agent collaborative prediction component capable of delivering reliable perception results even when connections to neighboring agents are inaccessible. Extensive experiments across several datasets demonstrate that CollaMamba outperforms state-of-the-art methods, achieving higher model accuracy while reducing computational and communication overhead by up to 71.9% and 98.2%, respectively. This work pioneers the exploration of the Mamba’s potential in collaborative perception.