The role of the covariance matrix is to represent the uncer-
tainty, to first order, in all the quantities in the state vector.
Feature estimates Ý can be freely added to or deleted from
the map as required, Ü and È growing or shrinking dynam-
ically. In normal operation, Ü and È change in two steps:
1. during motion, a prediction step uses a motion model
to calculate how the robot (or camera) moves and how its
position uncertainty increases; 2. when feature measure-
ments are obtained, a measurement model describes how
map and robot uncertainty can be reduced.
The critical importance of maintaining a full covariance
matrix È, complete with off-diagonal elements, has been ir-
refutably proven in SLAM research. These elements repre-
sent the correlation between estimates which is always in-
herent in map-building. The typical situation is that clusters
of close features will have position estimates which are un-
certain in the world reference frame but highly correlated
with one another — their relative positions are well known.
Holding correlation information means that measurements
of any one of this cluster correctly affects the estimates of
the others, and is the key to being able to re-visit and recog-
nise known areas after periods of neglect.
Successful SLAM approaches have generally operated
using not vision but specialised sensors such as laser range-
finders, and in somewhat restricted conditions including
2D planar robot movement and/or mapping, known robot
control inputs and accurately-modelled dynamics. In vi-
sion, Davison and Murray [6] made early progress in full-
covariance mapping using active stereo and Davison and
Kita [5], in perhaps the first work on SLAM in full 3D, used
đang được dịch, vui lòng đợi..