本文為PCL官方教程的Registration模塊的中文簡介版。
An Overview of Pairwise Registration
點雲配准包括以下步驟：
- from a set of points, identify interest points (i.e., keypoints) that best represent the scene in both datasets;
- at each keypoint, compute a feature descriptor;
- from the set of feature descriptors together with their XYZ positions in the two datasets, estimate a set of correspondences, based on the similarities between features and positions;
- given that the data is assumed to be noisy, not all correspondences are valid, so reject those bad correspondences that contribute negatively to the registration process;
- from the remaining set of good correspondences, estimate a motion transformation.
針對上述每一個步驟，PCL的registration模塊提供了多種算法進行實現 。
Keypoint
諸如 NARF, SIFT and FAST。
Feature descriptors
諸如NARF, FPFH, BRIEF or SIFT。
Correspondences Estimation
point matching
-
brute force matching,
-
kd-tree nearest neighbor search (FLANN),
-
searching in the image space of organized data, and
-
searching in the index space of organized data.
feature matching
-
brute force matching and
-
kd-tree nearest neighbor search (FLANN).
Corresdondences Rejection
使用RANSAC，或者剪出多余數據。
Transformation Estimation
諸如 SVD for motion estimate; - Levenberg-Marquardt with different kernels for motion estimate。
算法案例
其中（1）和（2）是point matching，（3）是feature matching。
（1）ICP
ICP的使用SVD求解轉換矩陣，其參考文章：
Least-Squares Estimation of Transformation Parameters Between Two Point Patterns
（2）NDT
參考論文：
1. The Three-Dimensional Normal-Distributions Transform an Efficient Representation for Registration, Surface Analysis, and Loop Detection. MARTIN MAGNUSSON doctoral dissertation。
2. Line Search Algorithm with Guaranteed Sufficient Decrease. 計算迭代步長。
（3）改進版RANSAC
參考論文：
Pose Estimation using Local Structure-Specific Shape and Appearance Context. ICRA 2013.
相關資料：