融合深度特征一致性与注意力网络的点云配准方法OA

Traditional point cloud registration algorithms are often observed to converge to local optima.This occurs when initial pose errors,noise,and repeated structures exist.To address this issue,a registra-tion method was proposed to integrate deep feature consistency constraints with an attention mechanism.An attention-enhanced deep feature extraction network(AENet)was constructed.It was trained via self-supervised learning to generate point-wise descriptors invariant to rigid transformations.Using these de-scriptors,initial correspondences were established.A coarse transformation was then estimated,provid-ing a reliable initialization for subsequent refinement.A deep feature consistency term was embedded into a multi-scale iterative closest point(ICP)optimization framework.It formed a joint objective that unified geometric alignment and deep feature matching for coarse-to-fine registration.By incorporating the feature similarity constraint into geometric optimization,a unified model was established.This model jointly en-forced geometric proximity and deep feature consistency throughout the registration process.Refinement was performed progressively.It proceeded through coarse,intermediate,and fine scales.This improved convergence stability and reduced sensitivity to challenging conditions such as large pose variations and structural ambiguities.Extensive experiments are conducted on the ModelNet40 dataset.Results demon-strate significant improvements across multiple error metrics.Specifically,the root mean square error of rotation(RMSE R)is reduced by approximately 85.5%,89.7%,78.8%,74.3%and 61.6%compared to FINet,OGMM,IDAM GNN,PREDATOR and RoCNet respectively.Similarly,the root mean square error of translation(RMSE t)is lowered by about 88.2%,87.0%,51.1%,72.0%and 18.2%.These results indicate that the proposed framework effectively improves both accuracy and robustness.It provides a practical solution for high-precision point cloud alignment.The method performs well under complex environments and structural ambiguities.