不确定性感知的RGB图像6D物体位姿估计方法OA

To address the limited accuracy of 6D object pose estimation under complex backgrounds and weak-texture conditions,an uncertainty-aware 6D object pose estimation framework was proposed.First-ly,a multi-scale receptive field enhancement module was introduced into the encoder-decoder architec-ture at the feature extraction stage.By integrating contextual information with different convolutional scales and dilation rates,the network was able to effectively improve the multi-scale feature representation capability for weak-texture objects.Secondly,a pixel-wise confidence-aware voting mechanism was de-signed to model the reliability of the predicted dense keypoint unit vector fields,and the contribution of each pixel was adaptively weighted during the voting process,thereby enhancing the robustness and accura-cy of keypoint localization.Finally,the uncertainty of keypoint localization was explicitly incorporated into the pose estimation stage by constructing an uncertainty-aware PnP framework,which effectively reduced the influence of noisy keypoints on pose estimation results.Experimental results on the LineMOD,Occlu-sion LineMOD,and YCB-Video datasets demonstrate that,in comparison with several representative methods such as BB8,DenseFusion,and PoseCNN,the proposed method achieves significant improve-ments in pose estimation accuracy,with an average increase of approximately 7％in the Average Distance(ADD)metric,and exhibits strong robustness under complex background and weak-texture scenarios.These results demonstrate the effectiveness,robustness,and adaptability of the proposed method under complex background and weak-texture conditions.