Page 180 - Introduction to Autonomous Mobile Robots

165
Perception
direction of illumination, the defocusing caused by optics, the side effects imposed by
nearby objects with different colors, and so on. Therefore the problem of visual feature
extraction is largely one of removing the majority of irrelevant information in an image so
that the remaining information unambiguously describes specific features in the environ-
ment.
We divide vision-based feature extraction methods into two classes based on their spa-
tial extent. Spatially localized features are those features found in subregions of one or
more images, corresponding to specific locations in the physical world. Whole-image fea-
tures are those features that are functions of the entire image or set of images, correspond-
ing to a large visually connected area in the physical world.
Before continuing it is important to note that all vision-based sensors supply images
with such a significant amount of noise that a first step usually consists of “cleaning” the
image before launching any feature extraction algorithm. Therefore, we first describe the
process of initial image filtering, or preprocessing.

Image preprocessing. Many image-processing algorithms make use of the second deriv-
ative of the image intensity. Indeed, the Laplacian of Gaussian method we studied in sec-
tion 4.1.8.2 for stereo ranging is such an example. Because of the susceptibility of such
high-order derivative algorithms to changes in illumination in the basic signal, it is impor-
tant to smooth the signal so that changes in intensity are due to real changes in the luminos-
ity of objects in the scene rather than random variations due to imaging noise. A standard
approach is convolution with a Gaussian distribution function, as we described earlier in
section 4.1.8.2:

ˆ G ⊗
I = I (4.82)
Of course, when approximated by a discrete kernel, such as a 3 x 3 table, the result is
essentially local, weighted averaging:

121
1
G = ----- - 242 (4.83)
16
121

Such a low-pass filter effectively removes high-frequency noise, and this in turn causes
the first derivative and especially the second derivative of intensity to be far more stable.
Because of the importance of gradients and derivatives to image processing, such Gaussian
smoothing preprocessing is a popular first step of virtually all computer vision algorithms.