3D point cloud course (VII) -- feature point description

1. What is characteristic point

1.1 Image feature points

ORB slam 

 

1.2 Point cloud feature points  

Point cloud registration ：ICP It is required to have a good enough initial translation and rotation matrix , And there is a certain coincidence rate

2. How to extract feature points

2.1 Image extraction feature points

2.1.1 Harris

  A good feature , The interior will change as it moves

U,V The smaller, the more sensitive   The characteristic point is x、y There is a large reciprocal in all directions

 

NMS Operate to filter feature points  

 

  The core idea ： After a small square moves , Inside Intensity Transform to select feature points . It turns into finding the covariance matrix in the block M, First order inversion in each direction .

2.2  Point cloud feature points

2.2.1 3DHarris

 

2.2.2 PCA 

Points with many points in three directions are characteristic points  

Traditional methods are very sensitive to noise , Can't use  

2.3 Deep learning feature extraction

2.3.1 USIP 

Unsupervised learning ： 1、 Feature points have nothing to do with point cloud rotation 2、 Feature points are related to scale It limits the perception domain

The previous column has high confidence , The next column is all feature points  

3. Description of feature points

3.1 Based on histogram

3.1.1 Histogram based

  Only care about the distance between points , Don't care about point distribution

3.1.2 Signature based

Rotate the same thing , The description will change  

3.1.3 PFH

The line between each point and the surrounding points

1、 Set up a coordinate system

2、 Calculating characteristics

Describe the change of points around feature points

3、 Each characteristic parameter establishes a histogram  

3.1.4 SPFH

Only the line between the feature points and the surrounding points is considered , Make three histograms

3.2 Based on coordinate system

3.2.1 SHOT

1. Set up a coordinate system LRF

2. Divide the space around the feature points into 32 block

3. Calculate the histogram of each small space , The length of each histogram is 11

There is a hard cutting problem , Unstable to noise , So soft cutting is proposed . Voting without direct bonus is either black or white , But the probability of becoming linear interpolation .

• summary

4. Feature point matching