Multi sensor fusion positioning (II) -- map based positioning

Catalog

One 、 Loop back detection  

1.1 be based on Scan Context

1.2 Based on histogram

One 、 Loop back detection  

Loop detection can only eliminate part of the error , All errors cannot be eliminated . It is relatively simple to use feature point descriptors in vision .

1.1 be based on Scan Context

Three dimensional to two dimensional , Match with image .

1、 Grid

 2、 Generate scan contest

3、 matching

The car faces close , The two figures are similar

In order to solve the big difference caused by different vehicle orientations , The image needs to be cut and translated to make the orientation similar

The author merges each column , Two dimensional to one dimensional , Find out where to cut roughly , Up to 3D fine matching

4、 Time complexity

How to quickly find matching frames

Count whether there are points in each grid , If the environment is similar , The number of points is similar .

1.2 Based on histogram

This is a solid-state lidar  

Two 、 Back end optimization

Add the a priori position and attitude of loop and inertial navigation , Correct the error of odometer .

2.1 Lie group lie algebra

Using convex optimization to build graphs will use lie group lie algebra . Lie groups are matrices , Lie algebra is a vector

 

2.2  Pose correction based on loop

2.2.1 Calculate the residual term  

2.2.2 Solve the optimization direction

​​​​​​

 

 2.3 A priori based optimization

3、 ... and 、 Establishment of point cloud map  

3.1 technological process

Put the odometer 、 Loop 、 A priori fusion .

In the first frame, inertial navigation and odometer have relative positions , Transform odometer multiplication matrix into inertial navigation coordinate system .

When the time difference is greater than a certain value, the closed loop is detected

3.2 Point cloud distortion

Calculate time through geometric relationship  

 

Four 、 Map based positioning

4.1 Overall process

In map matching , Robustness and running speed are more important , So in practice , be based on NDT The matching of is more widely used extensive .

 

Initialization problem