- corners. The corner detector implemented here is our
own version of the Plessey algorithm developed by Harris and Stevens.
This involves the generation of a corner strenth image, followed by
maxima detection and peak fitting to obtain the corner locations to
sub-pixel accuracy. Although detected corners do not always
correspond to intersection vertices identified by edge fitting
they still provide accurate positional information at reproduceable
positions in the image from three dimensional objects. Other
corner processors may give locations which are more consistent with
edge detection processes.
sigma : The parameter of the Gaussian convolution profile
used to filter the image prior to corner detection.
This effectively sets the minimum scale of detected
features.
precision : The ratio of the smallest stored value of the Gaussian
convolution profile.
- rectify. Use the current stereo camera geometry to compute
the rectified locations of each corner. This process must be performed prior
to stereo matching and geometry recovery. Rectified locations are added
as extra data to the existing structure so that the original location
of each corner is preserved. There is thus no need for the equivalent
derectification process in the Edge Tool.
- stereo. A correlation based matching algorithm that matches
corner features which are consitent with the current parallel camera
geometry. As this method is correlation based there is a fundamental
requirement that the local image patch surrounding each corner is
similar. Thus corners on extremal boundaries will normally not be sucessfully
matched.
Stereo Parameters:
lowdisp : The lower limit of the allowed disparity range in
fractions of image width from the centre of each
region of interest.
updisp : The upper limit of the disparity range.
width : The epi-polar band width used to define the valid
matching area (set large for poor camera geometry).
correlation : The minimum cross correlation measure required between
corners for them to be regarded as candidate matches.
uniqueness : The minimum difference in correlation values between
candidate matches before a match is labelled ambiguous.
This parameter is particularly important in regions of
many similar features where the likelihood of getting a
missmatch is high.
- temporal. As corner features are well located in both directions
on the image plane they can be matched between temporal frames with
relatively weak assumptions about the relative transformations between images.
Temporal Parameters:
height : Maximum vertical shift allowed between successive
locations of a corner in the image plane relative to
the centre of the region of interest.
width : Maximum horizontal shift.
- geom. Compute the 3D location of uniquely matched stereo corners
and make them available for display in the threed Tv.
