C14 - SfM SLAM

• structure from Motion

input:

• images with feature correspodences

output:

• structure: 3D positions of features

• motion: camera parameters including rotation and translation

  • intrinsic parameters optional

• hierarchical SfM

• Sequential SfM

• disorderes images

• estimated 3D structure and camera poses in bottom up way

• takes sequential imagesas input -> equivalent to visual odometry

• it incrementally estimates camera pose and 3D structure

• have set of disordered images

• -> independently detect features in them

• match features between the images (i.e. bruteforce; there exist more efficient methods that will not be introduced)

• build clusters from close frames (i.e. where we have feature matches…)

• generate topological tree based on number of matches

• start from terminal nodes to perform two-view SfM

• generate local model based on 2D-2D geometry and 3D-2D geometry (as introduced in chapter 9 or so…)

• merge different models based on 3D-3D geometry (-> point cloud registration)

• then jointly optimize camera poses and 3D points -> i.e. bundle adjustment using nonlinear least squares

• same as introduced in chapter 10

• -> Visual odometry (i.e. 2D-2D -> triangulate 3D -> find 3D-2D -> triangulate new 3D -> …)

• after triangulation to incremetn the map

• -> we perform bundle adjustment…

• simultaneous localization and mapping

• -> can be treated as combination of visual odometry and loop closure

• VO prone to be affcected by noise

• -> thus leads to drift error over time

• -> SLAM guarantees global consistency based on loop closure

• loop detection

• loop correction