Image Processing Reference
In-Depth Information
Therefore, the processing of the information obtained from the sensors has to
be quick and efficient in order to provide information in time, even if its quality is
reduced. Many research teams are working on architectures dedicated to this type of
application.
As for the methods that have been developed, several solutions can be suggested.
Data extrapolation. When the data is acquired, it is dated, either by a clock com-
mon to all of the sensors, or by different clocks that are synchronized periodically.
Based on a sequence of data and using an evolutionary model, also called a dynamic
model, it is often possible to extrapolate the observed variable, even if its quality
deteriorates over time, from lack of observations. This is typically what is done with
tracking filters, such as the Kalman filter [BAR 88, KOK 94, STE 98]. The prediction
can be used to obtain an element of information with an error greater than that of the
initial data, but which corresponds to the current date. This method is described at
length in section 2.2.1.
Focusing. Depending on the context, we might need to obtain certain elements of
information that seem more important than others. We would then have to focus the
system's perception abilities on these elements, even if that means overlooking others
that may also be available. For example, when a moving robot is about to go through a
doorway, the areas where the two jambs are located are observed in priority, even if a
watch is maintained to detect possible obstacles in front of the vehicle. The technique
is used in particular in the case of information in the form of an image, where only
a part of the image is subjected to significant processing. Focusing requires being
able to identify what is relevant based on the estimate of the situation, or on action
objectives that have been defined. Depending on what is relevant, a perception strategy
will then have to be defined in terms of the sensors used, the choice of the data and
the algorithms that are to be used. This method has been widely used at the LASMEA
laboratory in Clermont-Ferrand for aiming, based on the results of image analysis,
a laser beam capable of measuring the distance between two automobiles [CHE 96,
TRA 93]. Finally, the technique is the same as that presented for the identification of
military targets described in section 2.2.3.
Sensor based control. This method consists of foregoing the interpretation of data,
which is often very costly in terms of time and computing power, by computing the
control directly from the information obtained from the sensors before they undergo
any extensive processing. Depending on the robot's situation and the mission it has to
achieve, a control strategy is defined based on the sensor's measurements. For exam-
ple, if a moving robot is located in a hallway and has to maneuver between the two
walls that confine it, we can impose the order that the distances between each of the
walls and the robot have to be identical. The robot's movements are then directly cal-
culated based on the difference in the measurements by the sensors that provide these
two distances. This method can help substantially in speeding up the control loop, but
Search WWH ::




Custom Search