ORB(Oriented FAST and Rotated BRIEF)

   Always before ORB-SLAM Heard in various versions of ORB, But I don't know what is ORB. Take this opportunity to learn about ORB.
  ORB(Oriented FAST and Rotated BRIEF). First of all, realize that ORB Not the name of the feature , It is a method to complete feature matching between images . The matching step is reflected in the naming . Here are two questions ：

  1.ORB What are the advantages of ？

   It can solve two common problems of feature matching between images ：(1). Scaling problem (2). Rotation problem . Here is a detailed introduction ORB Two processes of , We can learn how to solve these two common problems

  2. Introduce ORB Two processes of

  ORB There are two characteristics ： Key points (Oriented FAST)、 Narrator (BRIEF, Binary Robust Independent Elementary Feature). The key point goes without saying , The reason why it is called descriptor is that it describes the image information around the key points , And use the description vector to express .
  ORB Contains two parts of improvement , Part of it is FAST, Improve to Oriented FAST. Part of it is BRIEF, Improve to Rotated BRIEF.

  ORB There are two steps ： feature extraction → Feature matching
  1） feature extraction
   Feature extraction includes two steps ：FAST Extract key points →BRIEF Descriptor calculation
    (1)FAST Corner point ( Key points ) extract
    FAST The determination steps of corner points are as follows ：

   Actually ORB The reason for the good effect lies in the effectiveness of feature extraction ： Divide the pictures into squares , Extract... Within the grid FAST Key points . Keep dividing ( Quadtree structure ), Until it approaches the preset value . Try to ensure that each grid has one FAST spot ( It is allowed that there is no FAST spot , Multiple cells are also allowed FAST spot , Multiple situations retain the largest response FAST Key points , To avoid corner concentration ), Finally, the distribution of feature key points is uniform . The feature points are evenly distributed , Ensure the effectiveness of feature extraction .
  Oriented FAST a FAST The improvement is to increase the rotation invariance 、 Scale invariance , After improvement, it can increase the robustness of representation in different images . The embodiment of rotation invariance is in addition to FAST The position of feature points in the also increases the direction . The resolution of scale invariance is based on image pyramid .
   Understanding of image pyramid ： A pair of pictures , Store with different scaling factors . According to the size of the picture to be matched , Select the appropriate level image for matching
   The above two improved strategies solve two problems of feature extraction ：(1). Scaling problem (2). Rotation problem
  (2)BRIEF Descriptor calculation
   Descriptor calculation is mainly to retain the information around key points , Used for matching between key points after .rBRIEF a BRIFT The rotation invariance is maintained , Just using the calculation of the direction in the corner calculation just now .

  2） Feature matching ：

   Feature matching is to find the corresponding matching points between two frames ( In image processing, the basic unit is frame , Data processing is performed in two adjacent frames ). The concept of feature matching is well understood , It is mainly the matching algorithm of feature matching . Here are three , But there are more than three
  (1) Violent match
   The main idea ： Calculate the descriptive sub distance of each pair of key points in two frames of images , Sort , Take the nearest as the matching point .
   The descriptor distance indicates the degree of similarity , The smaller, the more similar . The specific calculation process is no longer stated in this statement ( I don't know ~)
  (2)SearchForInitialization
   The main idea ： Delimit a certain area around the point to be matched , The histogram is established according to the direction of all other points in the area . Find the first three directions with the largest number of feature points , Determine the first main direction and the second main direction according to the judgment formula ( May not exist )
  (3)SearchByProjection
   The main idea ： The pose is estimated according to the uniform velocity model , Project the map corresponding to the point of the previous frame to the current frame , Calculate search points ( On current frame ) Hamming distance to the previous frame point ( Used for binary distance measurement ), If Hamming distance is less than threshold , It is considered to be its matching point . After posture optimization , The matching point satisfying the condition is greater than 10 It is considered that the two frame tracking is successful
   notes ： The binary corresponding bits are different , The distance between Han and Ming will increase 1. So the higher the degree of similarity , The smaller the distance between Han and Ming .

   This article is simple to ORB There's an impression ,ORB Itself is not my focus .ORB How to be in ORB-SLAM It is important to play a role in various versions of ( At present, it is not obvious ,ORB-SLAM In the description process, it seems that the specific description of the feature matching method is directly omitted , It just states the choice ORB As a feature matching method )

Reference material ：
1.ORB Feature matching
2.ORB Feature point matching
3.【 Detailed explanation 】ORB Feature extraction and matching _ Bili, Bili

Due to the limited level of the author , If something is wrong , Please correct in the comments area below , thank you ！