Some time ago, I studied Lecture 7 of Gao Bo's "Visual SLAM Fourteen Lectures" - Visual Odometer 1. This piece mainly introduces the feature point method in the front-end (visual odometer) part of visual SLAM. Among them, I compared the difference between the opencv library function and the handwritten ORB matching algorithm in my other blog. If you are interested, you can go and have a look. Welcome to comment and learn together. Links are provided below:
In this blog, I will not introduce too much about the related concepts of ORB feature point matching. I will mainly introduce how to estimate the camera motion based on the point pairs after getting the matched point pairs.
Table of contents
2.1.1 Solving process of epipolar geometry
2.1.2 Knowledge about homography matrix
2.1.3 Triangulation related knowledge
2.3.2 Nonlinear optimization method
The following is the framework of this part of the content I summarized:
1. Content framework
2. Main content
2.1 2D-2D: Epipolar geometry
For the knowledge about epipolar geometry, I made some study notes and summaries, as shown in the figure below:
2.1.1 Solving process of epipolar geometry
2.1.2 Knowledge about homography matrix
The following are notes related to the homography matrix:
2.1.3 Triangulation related knowledge
Here are the notes related to the triangulation:
2.2 3D-2D：PnP
For the 3D-2D situation, the PnP method is adopted. The following are my study notes:
2.3 3D-3D:ICP
2.3.1 SVD method
For the case of 3D-3D, the method used is ICP. The following are my study notes:
2.3.2 Nonlinear optimization method
The following are notes related to nonlinear optimization:
3. Summary
The above are my study notes on visual odometry 1 - feature point method. The writing is a bit scribbled. If you have any questions or suggestions, please leave a message in the comment area.