DF-CoopNet: Cooperative perception via local feature enhancement and global sparse attention

This is the authors' abstract. We don't add key points for this paper.

Cooperative perception technology plays a crucial role in autonomous driving systems by improving safety and enabling real-time decision-making. However, existing LiDAR point cloud processing methods face significant challenges in both local geometric feature extraction and global feature fusion. To address these issues, this paper proposes DF-CoopNet, a cooperative perception framework comprising two core modules: Local Geometry Enhancement (LGE) and Sparse Key Feature Attention (SKFA). The LGE module enhances local geometric representations using a deformable k-nearest neighbor graph structure and adaptive fusion mechanism to effectively detect occluded targets. The SKFA module introduces a hierarchical sparse attention mechanism that balances performance and computational complexity through a Top-k strategy. Extensive experiments on the OPV2V and V2V4Real datasets demonstrate that DF-CoopNet significantly outperforms existing methods while maintaining robust detection performance even with substantially reduced point cloud data, validating its effectiveness for real-world cooperative perception applications.