Overview of orb-slam3 paper

Compared with ORB-SLAM2,ORB-SLAM3 There are three major changes ： Uncouple camera model , The fusion IMU And a multi map system , Although as always, it is not an outstanding innovation , But the engineering test results are really good .

ORB-SLAM The system framework of the series has not changed much , The core threads of the system are mainly Trace thread , Local drawing thread , Loop detection thread .

（1） Trace thread ：
The input image （ and IMU data ）, Output the pose corresponding to the current frame .
（ It should be noted that , stay SLAM In the system , The output pose is the position and orientation information of the camera in the world coordinate system at each time , namely 3D Rotation and 3D The translation vector of . The origin of the world coordinate system generally defaults to the position of the camera corresponding to the successful initialization , That is, the pose of the first frame consists of a unit matrix rotation matrix and a displacement vector [0,0,0] form . The camera position of each frame calculated later is the rotation and translation transformation relationship of the first frame )

After the tracking thread receives the image information , By tracking the uniform velocity model , Tracking reference frames , Tracking local maps , Continuously optimize the pose of the current frame , After obtaining the current frame pose, the tracking thread decides whether to create a key frame according to the current system state .

in addition , The trace thread has relocation function , After the trace fails, it will pass bow Word bag vector search for scene recognition to quickly find the pose （ However, if a certain section of the road is always lack of features, then the location cannot be found , There is a high probability that you will not be able to find it when you lose it during real-time test in outdoor environment , At this time SLAM The system will be reinitialized . The difference between multiple maps is that if the tracking fails completely , It will keep the current map and participate in the subsequent loop detection , Each reinitialization starts from the same starting point , If loopback information is detected, the two sub maps with loopback will be fused ）.

（2） Local drawing thread ：
In essence SLAM Be divided into two parts ： Front end and back end .

The front end is the trace thread , The biggest purpose of the back-end including local mapping thread and loop back detection thread is to optimize the pose calculated by the front-end , With accurate posture SLAM The created map can have better consistency , because SLAM In essence, a map is a mosaic of images according to their position and pose , A sparse map is a mosaic of feature points , Using the whole image pixel mosaic is a dense map .

The local mapping thread will always maintain a small common view map near the current key frame （covisibility graph）, It consists of key frames and map points , And with SLAM The small map is constantly updated , Add or delete key frames and map points . After initialization with the system , The local mapping thread will always wait for and detect the input of the key frame of the tracking thread , Once a key frame is detected, the covisibility graph Conduct BA Optimize .

（3） Loop detection thread ：
stay SLAM Especially vision SLAM In the system , The pose of the camera is calculated according to the rotation and translation motion between images and the geometric projection relationship , In vision SLAM in , The pose at each frame time is in the previous frame （EKF） Or the previous several frames （ Nonlinear optimization ） Calculated on the basis of posture , In some big scenes , When SLAM When the system runs for a long time , The projection error of the camera will accumulate all the time , On the map, you will see that Mingming's camera has turned around and returned to the original point , But the map and the track are not connected .

Loop back detection is to receive key frames continuously Location recognition , adopt bow Word bag vector for fast similarity matching , If loop back is detected , Then the similarity transformation relationship between the key frames generating loop is established , According to the similarity constraint, the loop back key frame Essential Graph To optimize , After determining the optimized pose of the loopback key frame , Through propagation, the pose of the key frame that has a common view relationship with the loopback key frame is updated . The final result is the trajectory and map information with global consistency , Perfect integration of map information at the closed loop .