Page 145 - Dynamic Vision for Perception and Control of Motion

5.1 Visual Features 129

long-range, high-resolution viewing, the density of sensor elements is very high; in
the peripheral areas added for large angular viewing range, element density is
small. By this combination, a large viewing range can be combined with high reso-
lution at least in some area with relatively moderate overall data rates. The area of
high resolution can be shifted by active viewing direction (gaze) control by both
the eye and the head.
In technical systems, since inexpensive sensors are available only with homo-
geneous pixel distributions, an equivalent mapping system is achieved by mount-
ing two or mores cameras with lenses of different focal lengths fixed relative to
each other on a platform. The advantage of a suite of sensors covering the same
part of the scenery is that this part is immediately available to the system in a mul-
ti-scale data set. If the ratio of the focal lengths is four, the image produced by the
shorter focal length represents (coarse) information on the second pyramid level of
the image taken with the higher resolution (larger focal length). This dual scale fac-
tor may sometimes be advantageous in real-time vision where time delays are criti-
cal. On one hand, efficient handling of objects requires that a sufficiently large
number of pixels be available on each object for recognition and identification; on
the other hand, if there are too many pixels on a single object, image processing
becomes too involved and slow.
As mentioned in the introduction, in complex scenes with many objects or with
some objects with a complex pattern of sub-objects, relying solely on edge features
may lead to difficulties and ambiguities. Combining the interpretation of edge fea-
tures with area-based features (average intensity, color, or texture) often allows
easy disambiguation. Figure 5.3 shows a case of efficient real-time image sequence
processing. Large homogeneous areas can be tracked by both edge features and re-
gion-based features. In the near range, the boundaries between the regions are not
sharp but fuzzy (strongly perturbed, unsealed country road with grass spreading
onto the road). For initialization from a normal driving situation, searching edges
with large receptive fields in most likely areas is very efficient.
The area-based method covering the entire image width would improve robust-
ness to road parameters other than expected, but would also be costly because of

Figure 5.3. Combining edge and area-based features for robust object detection and rec-
ognition. Near range: Only edge detection in regions and with parameters selected accord-
ing to a normal driving situation. Far range: Four stripes covering the entire width of the
image; determine steep edges and intensity plateaus (lower part) to discover road forks.