Agriculture Reference
In-Depth Information
matches are always prone to appear, and the size of point clouds is often excessive
for real-time performance. In spite of activating several filters to reduce mismatches,
outliers sometimes end up forming part of the point clouds, leading to unrealistic
point locations. The subsequent analysis of these point clouds to extract key per-
ceptual features offers another opportunity to filter out spurious data and increase
the reliability of the system. The size of point clouds is proportional to the size of
the stereo images, which in technical terms means the image spatial resolution (
H
×
V
). The total number of pixels in the initial pair of stereo images is the maximum
number of correlated pixels that the matching algorithm can find, although in practi-
cal terms the set of correlated pixels will always be inferior to the image resolution
as a consequence of objects captured by only one of the two lenses, thus impeding
correlation between images, as well as due to filtered pixels with no stereo informa-
tion. The correlated pixels that pass the filters embedded in the matching algorithm
form the
disparity image
, a depth image from which 3-D coordinates are calcu-
lated. Although the point cloud of a moderate-size image can be easily managed, the
requirement of most vehicles to process data in real time poses a challenge to the 3-D
perception computer. A stereo camera can acquire visual information at 30 frames/s,
and control loops for automated vehicles should run at 10 Hz at least. An image of
resolution 640
480 has 307,200 pixels. If the disparity image is, for example, the
result of matching 70% of the pixels, the point cloud will consist of 215,040 points
that must be handled by the perception computer at the frequency demanded by
the vehicle. These difficulties for dealing with massive arrays of points have led
to several solutions to simplify data without sacrificing essential information. An
immediate step to ease the information behind point clouds is by projecting the data
into 2-D planes (frontal
XZ
, top
XY
, or side
YZ
) in combination with a quantization
of space through regular grids. What comes next is the methodology to compose the
grids by filling the cells, and it is here, however, where approaches differ based on
the objectives of the perception system. The concept of
3-D density
, and its practical
realization in
density grids
, has been effective for real-time obstacle detection and
crop mapping (Rovira-Más et al. 2006). The 3-D density is the number of stereo
correlated points per unit volume calculated for each cell of the grid. Although the
number of points is always related to the same volume of space, corrections need to
be introduced for two practical reasons. First, density values have to be normalized
for each image in such a way that obstacles possessing high values of density can
be identified with independence of the absolute magnitude of density, which usually
will vary from image to image as a consequence of changing illumination or texture.
Second, further objects in the field of view of the camera will necessarily be defined
by a smaller number of points than foreground objects because the field of view of
the camera opens with distance but the horizontal resolution of the image is constant.
Once these two corrections have been implemented, density grids can be compared
among them and global parameters applied to find obstacles, trajectories, or other
interesting features. When one—or both—of the stereo imaging sensors implements
color, the point cloud can be represented more realistically by assigning to each point
its corresponding RGB color code. Figure 12.5 shows the two ways described to deal
with stereo-based 3-D information for a vineyard scene: point clouds represented in
a 3-D Cartesian frame with ground coordinates (a), and the same scene simplified
×
