A Lightweight Two-Stage Multivehicle Feature Fusion Method Guided by Global Feature

Effectively selecting high-quality feature from each vehicle for fusion is crucial for collaborative perception performance. However, the uncertain differences in the perceptual ranges of each vehicle pose a challenge to fusion. Current fusion methods typically feed all received features into multi-layer self-attention to alleviate differences. This results in a high computational complexity that scales quadratically with connected vehicles number. To address this issue, we propose a lightweight two-stage multi-vehicle feature fusion method guided by global feature, which consists of two fusion stages. In the first fusion stage, we design a global feature builder (GFB) to effectively extract and aggregate the most representative features across all vehicles, constructing a feature map with a holistic perspective. Global feature can provide rich prior information for the next stage. In the second fusion stage, to effectively fuse reliable features and handle high computational complexity caused by stacking self attention layers, we introduce a global feature guided cross attention fusion (GGCF) module, which interacts global feature as query with features from each vehicle to obtain fusion weights. Importantly, we only need a layer of cross attention and the length of query is one, so our method has a computational complexity linearly proportional to connected vehicles number. To validate the effectiveness of our proposed method, we conduct experiments on simulated and real-world scenario datasets. Experimental results demonstrate that our fusion method has a faster inference speed than the state-of-the-art (SOTA) model, and the 3D detection performance achieves better performance in both real and noisy environments.