3D point cloud course (VIII) -- feature point matching

Catalog

1、ICP

2、ICP improvement  

2.1 Reduction point

2.2 Data Association

2.3 Outlier Rejection

2.4 Loss function

2、Normal Distribution Transform（NDT） 

2.1 NDT Dividing grid

2.2 MLE（Maximum Likehood Estimation）

3、 RANSAC

1、ICP

  First, find the corresponding problem between the two groups of point clouds , stay ICP Find the nearest point inside , Forcibly obtain R,t, It was obviously wrong at first , But this is an iterative process , until R,t The change is very small. . It can be used IMU Get initialized R,t.

Find out R,t Get the minimum value  

2、ICP improvement  

2.1 Reduction point

• Don't do all the points ICP
• Sampling under the point cloud
• Normal Space Sampling（NSS）

• Take characteristic points

2.2 Data Association

• Nearest neighbor search
• Normal shooting： Suitable for smooth point clouds

• Projection
• Feature descriptor matching

2.3 Outlier Rejection

Remove long distance

Remove mismatched points

2.4 Loss function

• Point-to-point
• point-to-plane

The distance from the point to the nearest point plane ,point-to-point May force the corresponding point ,point-to-plane Convergence is faster .LOAM

2、Normal Distribution Transform（NDT） 

2.1 NDT Dividing grid

You can use the information around each point , No need to do nearest neighbor search

Take one grid as a unit , Find the grid where the point falls , The information in the lattice is a Gaussian model to describe the distribution of point clouds .

Larger than 5 Only a point can Gaussian modeling  

Gaussian distribution of three dimensions  

2.2 MLE（Maximum Likehood Estimation）

Given the initial R,t, Determine which square the point falls into , Determine the Gaussian model .

The registration problem of point cloud becomes maximum likelihood estimation

outlier Points will cause the solution to collapse , Let the point probability be at least equal to a certain number

 

3、 RANSAC

In many cases, there is no initial solution , It's time to Feature detection + description + RANSAC

Feature point extraction 、 Description matches 、 iteration

General rigid point cloud registration