TY - JOUR
T1 - RU-SS2D
T2 - Pose-Robust State-Space Model for V2X Collaborative 3-D Object Detection
AU - Cao, Yitong
AU - Zhai, Di Hua
AU - Zhan, Yufeng
AU - Xia, Yuanqing
N1 - Publisher Copyright:
© 2005-2012 IEEE.
PY - 2026
Y1 - 2026
N2 - Accurate cross-agent alignment is fundamental to vehicle-to-everything (V2X) collaborative 3-D object detection, yet relative poses are often imperfect in practice, causing feature misalignment and degraded fusion quality. Existing methods for pose-robust collaborative perception often rely on explicit pose correction or multistage uncertainty estimation, which can introduce additional supervision requirements, computational overhead, and system complexity. In this work, we propose robust and uncertainty-aware 2D-selective-scan (RU-SS2D), a robust single-stage fusion framework for V2X collaborative 3-D object detection under pose errors. Built on the SS2D state-space model for efficient cross-agent bird's-eye-view (BEV) fusion, RU-SS2D combines pixelwise robust weighting with agentwise reliability weighting to suppress locally misaligned features and reduce the influence of globally unreliable collaborators. We further introduce a geometry-aware dynamic scale routing network to improve multiscale BEV feature representation by adaptively modeling location-dependent scale preferences prior to collaborative fusion. Experiments on OPV2V, V2X-Sim 2.0, and DAIR-V2X show that RU-SS2D consistently improves collaborative 3-D detection accuracy under clean poses as well as under both known and unknown pose noise patterns. It delivers stronger robustness than representative baselines while maintaining a lightweight design, with approximately 39% fewer giga floating-point operations (GFLOPs) than multistage methods and 69% fewer GFLOPs than a representative Transformer-based method.
AB - Accurate cross-agent alignment is fundamental to vehicle-to-everything (V2X) collaborative 3-D object detection, yet relative poses are often imperfect in practice, causing feature misalignment and degraded fusion quality. Existing methods for pose-robust collaborative perception often rely on explicit pose correction or multistage uncertainty estimation, which can introduce additional supervision requirements, computational overhead, and system complexity. In this work, we propose robust and uncertainty-aware 2D-selective-scan (RU-SS2D), a robust single-stage fusion framework for V2X collaborative 3-D object detection under pose errors. Built on the SS2D state-space model for efficient cross-agent bird's-eye-view (BEV) fusion, RU-SS2D combines pixelwise robust weighting with agentwise reliability weighting to suppress locally misaligned features and reduce the influence of globally unreliable collaborators. We further introduce a geometry-aware dynamic scale routing network to improve multiscale BEV feature representation by adaptively modeling location-dependent scale preferences prior to collaborative fusion. Experiments on OPV2V, V2X-Sim 2.0, and DAIR-V2X show that RU-SS2D consistently improves collaborative 3-D detection accuracy under clean poses as well as under both known and unknown pose noise patterns. It delivers stronger robustness than representative baselines while maintaining a lightweight design, with approximately 39% fewer giga floating-point operations (GFLOPs) than multistage methods and 69% fewer GFLOPs than a representative Transformer-based method.
KW - 3-D object detection
KW - collaborative perception
KW - vehicle-to-everything (V2X)
UR - https://www.scopus.com/pages/publications/105052405466
U2 - 10.1109/TII.2026.3728772
DO - 10.1109/TII.2026.3728772
M3 - Article
AN - SCOPUS:105052405466
SN - 1551-3203
JO - IEEE Transactions on Industrial Informatics
JF - IEEE Transactions on Industrial Informatics
ER -