The ORCA VII Vision System
The vision system receives a color raster image from the Orca's
downwardpointing camera at periodic intervals. The goal is to analyze
the image and determine the existence, position, and orientation of
any rectangular bins in the field of view.
The image is first transformed from its original colorspace and
resolution into a grayscale raster consisting of a more manageable
number of pixels. The colorspace transformation is chosen to emphasize
the white sixinch rims around the bins. The edges of this white
region, especially those that bound the dark bin interiors, are likely
to be by far the sharpest, straightest, and generally most significant
edges in the image. Furthermore, they are all mutually parallel or
perpendicular, so that once an overall axis orientation is chosen,
only two edge orientations (parallel to the two axes) are of interest.
Filters are first applied to the source image (a) to determine the x
and y directional derivatives of the image considered as an intensity
function of x and y. The filters are convolution kernels calculated by
taking the x and yderivatives of a twodimensional Gaussian. The
kernels are xy separable, meaning that each can be applied using a
series of two onedimensional convolutions (for a speed boost), and
they are steerable, meaning that they are a basis for a set of
directional derivative filters for all directions that can be obtained
by linear combination. The results of applying the two derivative
filters are two signed rasters (b) holding the derivatives for each
pixel. Together they define a gradient vector for each pixel.
Plotting these gradients (c) clearly reveals the dominant orientation
of the edges in the image. The software detects this orientation by
combining the vectors using two orthogonal quarterturn symmetric
weighting functions. Using this new information, the x and
yderivative rasters can be combined to yield two new orthogonal
directional derivative rasters (d), this time specifically emphasizing
edges in line with the detected orientation. The next step is to
detect the individual locations of the strongest edges, and for
simplicity the software works with mathematical lines rather than
segments. Since only a single line orientation is of interest for each
raster, the rasters can be collapsed along this direction into
onedimensional functions (e). The most prominent extrema of these
functions (found by looking for zerocrossings of their derivatives)
are recorded as edgeline candidates (f, red and blue).
It then remains to consider combinations of these lines as possibly
bounding bin interiors. A proper set of four edges must obey many
constraints, including edge sign (opposite edges must have opposite
sign), bin size (compared to expectations based on the Orca's other
sensors), bin aspect ratio, gradients across the edges, and brightness
and consistency of the bininterior region of the image. Bin
candidates are identified and ranked using a confidence measure based
on these criteria, and likely bins are reported.
