复杂关节物体六自由度位姿估计与实时系统OA

To address limitations in kinematic modeling for 6-DoF pose estimation of complex articulated objects,a joint-aware decoupled pose estimation framework is proposed.A decoupled modeling strategy is adopted,in which rotation and translation are regressed through independent branches,thereby reducing error coupling.An attention-guided multimodal feature fusion mechanism is developed to capture the rela-tionships between image semantics and point cloud geometry,enhancing robustness under occlusion.An adaptive weighting scheme is further introduced to balance multiple loss terms during training.In addition,a lightweight module is designed to predict part-level articulation parameters.Experimental results on the ArtImage dataset demonstrate that the proposed method achieves an inference speed of up to 20 ms per frame without requiring external optimization.For complex objects such as drawers,the rotation error re-mains stable at 1.6°,while the 3D IoU for the glasses category improves by approximately 28%.The method consistently outperforms baseline approaches across most categories and evaluation metrics.The proposed framework enables unified estimation from the base pose to articulated parts,improving both pose accuracy and inference efficiency while preserving physical consistency.