Unicycle-like Robots with Eye-in-Hand Monocular Cameras: From PBVS towards IBVS - Visual Servoing via Advanced Numerical Methods - page 333

assumed to be close enough so that a significant number of features remain in view
all along the maneuver. The purpose of this section is to define the tools that enables
its use to servo the vehicle in the large , i.e. across paths connecting totally different
initial and final views. The necessary information are stored in images, called way-
points , which can be used to topologically connect the initial and desired images,
andina metric map, which stores sufficient data to implement the PBVS. A repre-
sentation of the environment that conveys these metric and topological information
will be referred to as a hybrid visual map .
The literature on VSLAM is rather extensive (see e.g. [27, 10, 12, 29]), and COTS
software is already available ([19]). These results are clearly fundamental for the
approach to servoing here presented, which instead relates to nonholonomic vehi-
cles. In the hybrid map representation, the metric information is represented by a
set of robot postures, along with the corresponding 3D position estimates for the
features observed from such postures. The topological information is represented
by an undirected reachability graph (we assume indeed that possible environment
changes do not affect the traversability of the space by the robot, [14]). The hybrid
map construction method is described in what follows:
1. from the initial unknown position of the vehicle ( i.e. W 0] T )animage

I A of a portion of the scene in view is grabbed and stored in the first node A of
the hybrid map (see Figure 18.2(a)));
2. from the image in view, a subset F A of n A features is selected;
3. the vehicle moves, avoiding obstacles with proximity sensors, in an arbitrary
direction using a simple control law that keeps the image point features in view;
4. an EKF is implemented using odometry and camera measurements to estimate
the relative spatial position of the feature in camera frame

ξ A = W [0 , 0 ,

;
5. once 3D feature position estimates have converged to a value under given level
of uncertainty determined by the covariance matrix, the robot stops moving,
updates the metric map (Figure 18.2(b)) and then a new node corresponding to
the current pose is added to the hybrid map;
6. to add new nodes from the already created ones, the procedure starts again from
step 2.

C

ξ 1 , ξ 2 , ξ 3 ] T (or a
node A with image I A ). Suppose that the robot has to reach a new position, say

W

ξ A = W [ Let the robot be in a generic mapped position, say

ξ K ,

expressed in the metric map. If W K corresponds to a topologically mapped location

K , which has an associated image I K (found using SIFT technique), a standard graph
visiting algorithm is used for the path selection from A to K in the image map,
therefore, permitting the vehicle to steer through the map nodes using the servoing
presented previously. For space limits, we refer the interested readers to [14] for
further details on the mapping/navigation processes here briefly discussed.

ξ

18.3.5 Experimental Results

Two experiments, with different experimental set-up are reported. The first exper-
imental setup comprises of a TRC LabMate vehicle, equipped with an analogical