Robust Multi-Agent Collaborative Perception via Triple-Attention and Dynamic Gating

Multi-agent collaborative perception can mitigate individual blind spots by sharing and integrating sensory data, thereby providing a more comprehensive understanding of the surrounding environment. However, challenges remain, particularly in feature fusion, where issues such as redundancy, localization errors, and the trade-off between perceptual accuracy and real-time performance must be addressed. In this paper, we propose the Triple-Attention Collaborative Perception (TA-CoPe) framework, which employs hybrid channel attention, frequency attention, and spatial attention to enhance feature representation across these dimensions, optimizing the fusion and selection of multi-dimensional features. Additionally, a dynamic gating mechanism with adaptive weight learning assigns varying weights to each module, dynamically optimizing their contributions to the final output, thus improving both the accuracy and robustness of feature fusion. Experimental results demonstrate that the proposed method significantly outperforms existing state-of-the-art methods on two public datasets in multi-agent perception tasks.