基于特征相似性的RGBD点云配准

三维点云数据的配准是计算机视觉领域的重要研究课题，也是三维重建的关键步 骤。针对 RGBD 点云数据的配准问题，提出一种基于特征相似性的初始配准方法。首先需要计 算待配准的 RGBD 点云模型的曲率和颜色特征度(CFD)，并对 CFD 进行统计分析，若模型颜色 特征足够丰富优先采用颜色相似性策略，反之尝试曲率相似性策略。通过特征点提取精简点云 模型，利用确定的对应点选择策略选择候选对应点对。在候选对应点对上采用优化样本一致性 算法获得初始配准变换矩阵，实现两片点云的初始配准。针对不同颜色纹理的 RGBD 点云模型， 本文方法可以自适应选择合适的特征点选择策略，实现点云间良好的初始配准。实验结果表明， 对于几何特征不明显的 RGBD 模型，本文方法能够自适应选择颜色相似性策略来较好地完成初 始配准。对于不同类型的模型配准结果较好，算法效率更高。

RGBD Point Cloud Registration Based on Feature Similarity

The registration of 3D point cloud data is an important research topic in the field of computer vision and a key step in 3D reconstruction. Aiming at the registration problem of RGBD point cloud data, a coarse registration method based on feature similarity is proposed. Firstly, the curvature and color characteristics of the RGBD point cloud model to be registered should be calculated. Through the statistical analysis of color characteristics, if the color features of the model are rich enough, the color similarity strategy will be adopted first, otherwise, the curvature similarity strategy will be tried. The feature point extraction can simplify the point cloud model. And we will use the corresponding point selection strategy to select all corresponding point pairs. The coarse registration matrix is obtained by adopting the optimized sample consensus algorithm on the candidate corresponding pairs, and the coarse registration of the two point clouds is realized. For the RGBD point cloud model with different colors and texture, this method can adaptively select the appropriate feature point selection strategy to realize the good coarse registration between point clouds. For different models, we can adaptively select the corresponding selection strategy to calculate the transformation matrix and complete the coarse registration. The experimental results show that the proposed method can adaptively select the color similarity strategy to complete the coarse registration for the RGBD model with less geometric features. For different types of model, the registration results are better, and the algorithm is more efficient.