Pose-aware BEV feature refinement for robust cooperative perception under pose uncertainty

Abstract In Vehicle-to-everything (V2X)-enabled autonomous driving systems, collaborative perception allows multiple agents to exchange complementary information, thereby enlarging the perceptual range and alleviating occlusion. However, in real-world deployments, pose estimation errors and measurement uncertainties can undermine cross-agent feature consistency. Even after global geometric alignment, the fused bird’s-eye-view (BEV) representation may still suffer from residual local spatial misalignment, contextual inconsistency, and unstable feature responses, which degrade the robustness and localization accuracy of collaborative 3D object detection under complex perturbations. To address this issue, we propose a pose-aware BEV feature refinement method for post-fusion BEV representations. The proposed method progressively improves fused BEV features through three stages: local spatial compensation, multi-resolution contextual enhancement, and reliability-aware response filtering. In this way, residual post-fusion errors can be further suppressed without introducing additional supervision. Built upon the OpenCOOD framework and the PointPillars detector, the proposed method is evaluated on the DAIR-V2X dataset under both noise-free and perturbed settings with Gaussian and Laplacian pose noise. Compared with the strong baseline CoAlign, it improves AP@0.5 and AP@0.7 by 3.5 and 3.8 percentage points, respectively, under noise-free conditions. Under severe pose perturbations (0.6 m, 0.6°), it further improves AP@0.7 by 1.3 and 1.4 percentage points under Gaussian and Laplacian noise, respectively. Ablation studies and qualitative results further verify the effectiveness of the proposed design. These results demonstrate that refining post-fusion BEV features is an effective way to mitigate residual feature-level errors after global alignment and improve the detection robustness and localization accuracy of collaborative 3D object detection under pose uncertainty.