Fusing Monocular Information in Multicamera SLAM

Author: Joan Solà

Year: 2008

Notes:

each camera used as a mono sensor and all info is fused in a SLAM filter => desynchronisation of cam possible, self calibration, cooperative SLAM
mono cameras = BO (bearing only) sensor
difference between Sfm and SLAM: Sfm is solved offline with BA and SLAM is online
Sfm minimize error in measurement space, SLAM in state space
stereo: limited 3D range, mechanical constraint, measurement space (u,v,d) disparity->0 for a point at infinity
advanced VO: hardware based image processing, planar homographies (?), tightly coupled BA
fusion of laser + camera: data association pb
here two sensors of same kind: appearance based matching possible

3D estimability:

conic shaped pdf for the landmark position
a landmark depth is estimated if the region of intersection of the rays is closed and small
depth recovery impossible in the center when moving forward
when moving laterrally the best estimable region is located in the center
core algo = EKF SLAM based on inverse depth, each lmk represented by $$ i = [x_0, \theta, \psi, \rho] $$ $\rho$ is the inverse depth of the euclidean distance to $x_0$, the angles are in global frame and $x_0$ is the position of the camera at init time
as cameras are angular sensors vision measurements are more sensitive to camera rotation than translation

Multicamera:

self calibration by including the relative orientation of the second camera in the EKF
cooperative cameras: a master detect and initialize landmarks and a slave reobserves the map built by the master

Perception:

Use active search (=searching features with prior info) as image processing framework
strong Harris kp are selected and a 15*15 patch around it as a descriptor
d>0 => positive depth, d<0 => negative depth
distribution of inverse depth can be projected on the image => elliptic region in gaussian case
prior: infinite depth ie inverse depth
Affine patch warping for correlation matching to predict the best current landmark appearance

Experiments:

self calibration: estimation of euler angles between cameras -> converge to correct result
outdoor experiment for colaborative SLAM