Visual SLAM algorithms: a survey from 2010 to 2016 https://share.summari.com/articles101186s41074-017-0027-2?utm_source=Webapp
- Simultaneous Localization and Mapping (SLAM) is a technique for obtaining the 3D structure of an unknown environment
- This technique was originally proposed to achieve autonomous control of robots in robotics.
- Since the input of such SLAM is visual information only, the technique is specifically referred to as visual SLAM (vSLAM).
- Define a certain coordinate system for camera pose estimation and 3D reconstruction in an unknown environment
- In the initialization, the global coordinate system should first be defined, and a part of the environment is reconstructed as an initial map
- Tracking and mapping are performed to continuously estimate camera poses
- 2D-3D correspondences between the image and the map are first obtained from feature matching or feature tracking in the image
- Camera pose is computed from the correspondences by solving the Perspective-n-Point (PnP) problem
- Relocalization: When the tracking is failed due to fast camera motion or some disturbances, it is necessary to compute the camera pose with respect to the map again.
- Global map optimization: When a map is revisited such that a starting region is captured again after some camera movement, reference information that represents the accumulative error from the beginning to the present can be computed and a loop constraint from the reference information is used as a constraint to suppress the error in the global optimization.