Page 50 - Autonomous Mobile Robots

Visual Guidance for Autonomous Vehicles 33

One of the key problems in processing ladar data is data association.
For example, consider capturing data from a tree. The points that are detec-
ted depend on the viewpoint: that is, surface features are not pose invariant
[10]. This problem becomes easier with the use of a putative model of the tree
whose 2D position is determined by a centroid, which is invariant. Such a model
is easier to initiate if image data provides the evidence that the data points match
image features with the correct “tree-like” attributes. Once we have a model we
can anticipate where to search for features to match data points and vice-versa.
In this case we want to compare the real data with a model prediction but this
has to be very efficient given the large amount of data and hypotheses that will
occur. A typical problem is to test if a model patch will be detected by a sensor,
and how many data points to expect. Range detection is equivalent to ray inter-
section and is more easily solved after projection into a 2D space: a cylindrical
projection is sufficient and preserves the topology.
To summarize, in isolation there is much ambiguity in either sensor, and
exchanging information using image constraints can reduce this problem. The
difficulty is how to implement this practically as the concept of “being like
a tree” is more abstract than the neat formulation of raw data fusion as seen
in Section 1.3.6. This lack of precise mathematical formulation and reliance
on heuristic rules deters many researchers. However, recent advances and
increasedprocessingspeedshavemadeprobabilisticreasoningtechniquestract-
able and worthy of consideration in real-time problems such as visual guidance
and terrain assessment.

1.4 CHALLENGES AND SOLUTIONS
The earlier sections have detailed many of the practical difficulties of visually
based guidance and presented pragmatic techniques used during field demon-
strations. To be realistic, autonomous vehicles represent a highly complex set
of problems and current capability is still at the stage of the SAP/F “donkey”
engaged in A-to-B mobility. To extend this capability, researchers need to think
further along the technology road map [1] and tackle perception challenges
such as: terrain mapping, detection of cover, classification of vegetation, and
the like.

1.4.1 Terrain Classification
Obstacle detection based only on distance information is not sufficient. Long
grass or small bushes will also be detected as obstacles because of their height.
However, the vehicle could easily drive through these “soft” obstacles. Altern-
atively, soft vegetation can cover a dangerous slope but appear as a traversable
surface. To reduce unnecessary avoidance driving, detected obstacles need to

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 33 — #33