UECP: Uncertainty-Enhanced Collaborative Perception

In collaborative autonomous driving, agents share sensory information to improve perception. A major challenge is determining how much to trust each collaborator's data. Existing methods typically use confidence maps derived from the detector's own classification output, which creates a circular dependency: the fusion process is biased by the very detection noise it aims to suppress. This paper asks: can we create a physically grounded, unbiased metric for perception quality that improves multi-agent fusion?

The authors introduce the Uncertainty-Enhanced Collaborative Perception (UECP) framework. Instead of relying on model-derived confidence, they propose an uncertainty map supervised directly by raw LiDAR point density. Because physical factors like occlusion or surface reflectivity manifest as reduced point returns, point density serves as an objective, sensor-level proxy for reliability.

To integrate this, the authors developed the Uncertainty-Aware Pyramid Fusion (UAPF) module, which employs two key mechanisms:

• Uncertainty-Weighted Downsampling (UWD): A technique that uses the uncertainty map to weight feature contributions during downsampling, ensuring that high-quality information is preserved while noise is suppressed.
• Uncertainty-Guided Residual Fusion (UGRF): A fusion strategy that uses uncertainty-weighted summation to integrate neighbor features into the ego-agent's representation, using a residual design to stabilize training and prevent unreliable data from corrupting the ego-agent's local features.

Experiments on the DAIR-V2X and V2V4REAL datasets demonstrate that UECP consistently outperforms state-of-the-art collaborative perception methods. By decoupling the quality metric from the detection head, the model effectively mitigates false positives and false negatives. The authors show that the uncertainty map provides a more reliable signal for weighting collaborative inputs than traditional confidence maps, leading to superior robustness in noisy or challenging driving scenarios.