Agriculture Reference
In-Depth Information
12.3 THREE-DIMENSIONAL AWARENESS FOR
NAVIGATION AND 3-D MAPPING
The move from 2-D to 3-D is, in practical terms, much more consequential than the
mere addition of a coordinate or dimension; it provides the
range
or distance between
the perceiving sensor and its surrounding objects. This means that in addition to the
full positioning of detected objects in a Cartesian frame fixed to the vehicle, safe-
guarding distances are readily available without the need for rotating mirrors such as
lidar heads or intricate arrays of ultrasonic devices. Although lidar range finders may
actually yield 3-D maps (Yokota et al. 2004), the complexity of coordinating a laser
beam spinning simultaneously in two perpendicular planes while a vehicle traverses
off-road terrains has led to a constant decrease in interest by field roboticists in favor
of
binocular stereovision
, which has been growing as computer power increases and
the cost of electronics diminishes. For this reason, from now on in this chapter we
will only consider binocular stereoscopic cameras as providers of
3-D surrounding
awareness
.
The successful assemblage of a stereovision perception system depends on the
proper selection of camera parameters. In other words, given the estimated
ranges of
actuation
for a particular application, two fundamental parameters need to be opti-
mally determined: the
focal length
f
(mm) of both lenses, which has to be exactly the
same, and the
baseline
B
(mm) or horizontal separation of both lenses measured at
their optical center. Both parameters are coupled, and it is their combination in real-
ity that results in the preferred embodiment for covering a determined space around
a vehicle. As a result, different combinations of
B
-
f
may lead to the same level of
perception. However, it can be a tedious task to follow a trial-and-error approach,
and certain framework is necessary to start considering baselines and lenses to cover
a chosen interval of ranges. To assist in this task, Rovira-Más et al. (2009) defined
the
form factor
FF of Eq. 12.1, where
B
is the baseline (mm),
f
is the focal length of
the lenses, and
R
is the targeted range (mm). Experimental work demonstrated that
form factors between 1.2 and 2 result in the optimum perception for sensing in 3-D
around intelligent vehicles. Some stereo cameras feature interchangeable lenses and
adjustable baselines, but this option is not convenient for field applications because
every time the baseline changes or a new pair of lenses is mounted, the camera
needs to be calibrated, which is not always convenient because the camera may be
located at inaccessible places within the vehicle. Precalibrated cameras with fixed
baseline and lenses, on the other hand, offer many advantages for sensing in off-road
environments, but parameter selection is critical because no alterations can be made
once the stereo head has been constructed. Figure 12.3 provides an example of how
Eq. 12.1 can be applied to select the optimum interval of baselines according to the
expected ranges. As for the three key parameters (
R
,
f
,
B
), it is helpful to fix one
of them and analyze the behavior of the other two. In Figure 12.3, lenses of 8 mm
(habitual for sensing around a vehicle) have been selected, and therefore for each
baseline (abscissa axis) between 5 cm and 1 m, an interval of recommended ranges
can be read by just drawing a vertical line at the preferred baseline. Notice that as
baselines increase, the minimum range available also increases. The valid area of
Figure 12.3 will change for each focal length selected. The form factor FF can also
