User manual

Small Vision System User Manual 42
3.6 Filtering
Like most vision algorithms, the results of stereo processing can contain errors. In the case of stereo,
these errors result from noisy video signals, and from the difficulty of matching untextured or regularly
textured image areas. Figure 3-9 shows a typical disparity image produced by the SRI algorithm. Higher
disparities (closer objects) are indicated by brighter green (or white, if this paper is printed without color).
There are 64 possible levels of disparity; in the figure, the closest disparities are around 40, while the
furthest are about 5. Note the significant errors in the upper left and right portion of the image, where
uniform areas make it hard to estimate the disparity.
In Figure 3-9(c), the interest operator is applied as a postfilter. Areas with insufficient texture are
rejected as low confidence: they appear black in the picture. Although the interest operator requires a
threshold, it’s straightforward to set it based on noise present in the video input. Showing a blank gray
area to the imagers produces an interest level related only to the video noise; the threshold is set slightly
above that. Or, more simply, you can use the temporal variance of poorly textured matches to adjust the
texture threshold. Observing the disparity image during realtime display, there will usually be areas that
flicker rapidly. Adjust the threshold upward until these regions disappear. If there are no such regions,
adjust the threshold downward until just before they appear.
(c) Texture filter applied
(d) Left/right and texture filter applied
(a) Input grayscale image, one of a stereo pair (b) Disparity image from area correlation
Figure 3-9 Post-filters applied to a disparity image. (c) is a texture filter that eliminates
textureless areas. (d) is a consistency check between left and right stereo matches.
There are still errors in portions of the image with disparity discontinuities, such as the side of the
subject’s head. These errors are caused by overlapping the correlation window on areas with very different
disparities. Application of a left/right check can eliminate these errors, as in Figure 3-9(d). The left/right