基于特征点优选全局配准的头部三维重建方法

Head 3D Reconstruction Method Based on Selected Feature Point and Global Registration

摘要: 为实现高精度的头部三维重建，解决因重叠率低和累积误差大导致的模型变形或不完整的问题，文章提出了一种以特征点优化选择为核心的全局配准方法。预处理阶段，从初始点云数据中分割出干净的头部区域。特征提取阶段，先基于几何分布选择出高重叠区域，再结合双向KD树技术与动态阈值双约束策略，对高重叠区域的特征点集进行优选，从而提高匹配点对应关系的可靠性。最后在配准过程中，利用筛选出的可靠点对进行模型到帧优化的多帧全局配准，在每次迭代后及时更新模型，从而增强帧间重叠，并将累积误差降至最低，最终重建出完整的头部三维模型。基于多角度RGB-D数据进行消融实验与对比实验，结果证明了所提方法的有效性，重建模型的均方根误差(RMSE)为1.7623 mm，效果优于其他方法。该方法显著提升了多视角点云的配准精度，实现了精确性与完整性兼备的头部点云三维重建。

Abstract: To achieve high precision head 3D reconstruction, and address the issues of head model deformation or incompleteness caused by low overlap ratio and large cumulative errors, this paper proposes a global registration method centered on feature point optimization selection. During the preprocessing stage, the clean head region is segmented from the initial point cloud data. In the feature extraction stage, high-overlap regions are selected based on geometric distribution, and then the point sets within them are refined and selected using bidirectional KD-tree technology and a dynamic threshold dual-constraint strategy to enhance the reliability of point correspondences. Finally, during the registration process, the model-to-frame optimized global registration of multiple frames is based on the selected reliable point correspondences. The model is updated promptly after each iteration to enhance inter-frame overlap and minimize cumulative errors, ultimately reconstructing a complete 3D head model. Ablation and comparative experiments using multi-angle RGB-D data demonstrate the effectiveness of the proposed method, with the root mean square error (RMSE) of the reconstructed model being only 1.7623 mm, and its superiority over other methods. The method significantly improves the registration accuracy of multi-viewpoint point clouds and achieves the 3D head reconstruction with both high accuracy and integrity.