RAAlign: Multi-Agent Feature Alignment and Reliability-Aware Fusion for Vehicle–Infrastructure Cooperative Object Localization and Perception
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Existing V2X cooperative perception methods mainly rely on LiDAR-only or LiDAR-camera inputs,and remain vulnerable to adverse weather and cross-agent alignment errors. We introduce RAAlign, a reliability-aware LiDAR–4D Radar fusion framework for cooperative 3D object detection. RAAlign exchanges quality-gated complementary messages, promotes object-level consistency across agents and modalities using an InfoNCE objective , and compensates for residual offsets before weighting neighbor features using spatial and temporal reliability for attention aggregation. Experiments on V2X-R show that RAAlign improves average [email protected] from 69.05 to 72.99 on AttFuse and from 70.24 to 73.64 with MDD, with the largest weather gain reaching 2.24 percentage points in snow. These results provide evidence that reliability-aware cross-agent LiDAR–4D Radar fusion improves cooperative perception under adverse weather and imperfect alignment.