Many conventional systems are limited in that, in order to facilitate the initial search for matches between left and right images, their cameras either need to be aligned so that the epipolar lines are the scanlines of the images or they need to be precisely calibrated with respect to a common external calibration target in order to derive the epipolar geometry.
Recent work in the area of weak calibration shows that the epipolar geometry can be derived by matching features between the two images without the use of an external calibration target [4][3]. In practice, we use the program described in [17] in order to compute the epipolar geometry.
Moreover, the epipolar geometry can be used for computing a 2-D rectification transformation which brings the images into a standard configuration, irrespective of the physical configuration of the cameras. Once the rectification matrices are computed, rectifying an image is fast because the mapping between pixel positions in the original image and the corresponding positions in the rectified image is stored in a look-up table. The rectification of a full 480x640 image takes 0.2s on a Sparc-10 workstation.
Once both images are rectified, the best match to a pixel in the left image is
found by evaluating a normalized correlation criterion at the pixels in the
right image along the corresponding scanline. The search is limited to a
pre-set disparity interval .
The stereo algorithm is implemented as a random access function which takes as input a pixel position
and a disparity interval
, and outputs the corresponding disparity. We chose this implementation
over the conventional raster scan implementation because we need to compute only a small number of matches compared to the
total number of pixels in the image, as we will show in Section 2.3.
Furthermore, as we will show in Section 2.2, the anticipated range of disparity
can
be adjusted dynamically at each pixel.
It is well known
that any stereo matching algorithm is prone to false matches due to lack of texture, repetitive texture, and occlusions.
In our case, false matches are of great consequence because they directly impact the behavior of the entire driving system.
To reduce instances of false matches, we implemented three levels of
mismatch detection. First, pixels at which the variation of intensity in the neighborhood, , is
insufficient are not matched. This test eliminates points on uniform surfaces such as roads.
Second, matches with an insufficient correlation value
are not matched. This test eliminates points in occluded parts, for example. A
third test finds the local extrema of the
correlation curve
at each pixel and discards the pixel if there exists a local maximum
such that
, where
is the global maximum value of correlation.
The current values used in the three tests are:
,
, and
.