Page 252 - Innovations in Intelligent Machines

Toward Robot Perception through Omnidirectional Vision 245
Visual Path Following

Visual Path Following can be described in a simple manner as a trajectory
following problem, without having the trajectory explicitly represented in the
scene. The trajectory is only a data structure learnt from example / experience
or specified through a visual interface.
Visual Path Following combines the precise self-localisation (detailed in
the preceding sections) with a controller that generates the control signals for
moving the robot, such as that proposed by de Wit et al [21].
Experiments were conducted using an omnidirectional camera with a
spherical mirror profile (shown in Fig. 3), mounted on a TRC labmate mobile
robot. Figure 8 illustrates tracking and self-localization while traversing a door
from the corridor into a room. The tracked features (shown as black circles)
are defined by vertical and ground-plane segments, tracked in bird’s eye view
images.
Currently, the user initializes the relevant features to track. To detect the
loss of tracking during operation, the process is continuously self-evaluated
by the robot, based on gradient intensities obtained within specified areas
around the landmark edges (Eq. 18). If these gradients decrease significantly
compared to those expected, a recovery mechanism is launched.

Fig. 8. Feature tracking at three instants (a,b,c), scene-model tracking in the robot
coordinate system (d) and the self-localisation result obtained by fixing the tracked
scene-model (e)