The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
863
In Table 6, |a| xyz , |a| xy , and |a| z are magnitudes of the
acceleration vector during a single step in 3D, horizontal, and
down directions, respectively; Var(|a| xy2 ), Var( |a| xy ), and
Var(|a| z ) are the corresponding variance of the acceleration
vector; Max(|a|) and Min(|a|) are the maximum and minimum
values of the acceleration for each pace.
An alternative implementation of the SL/SD
calibration/prediction module is based on FL (see, Moafipoor et
al., 2007a for details of this algorithm). By incorporating Fuzzy
Logic to our KBS, better process control is facilitated, as this
approach allows an easy addition of constraints, such as, for
example hallway layout for indoor navigation, or digital map
information, which are difficult to handle in “regular” EKF
environment. Fuzzy Logic can be described simply as
“computing with words rather than numbers,” and Fuzzy Logic
control can be described as “control with sentences rather than
equations” (Sasiadek and Khe, 2001). Rule-based Fuzzy Logic
provides a formal methodology for linguistic rules resulting
from reasoning and decision making with uncertain and
imprecise information. In fuzzy behavior-based navigation the
problem is decomposed into simpler tasks (independent
behaviors), and each behavior is composed of a set of Fuzzy
Logic rule statements aimed at achieving a well defined set of
objectives; example rules are:
Rule (/): If is AND is A^ ,...,AND
x im is A im THEN y is B i (!)
where i=l,...,n, and n is the number of rules in a given fuzzy
rule base; j=l,..,m, and m is the number of antecedents; x ;; are
V
the input variables, premise variables, which are the sensor data
of the mobile user; Ay are the input fuzzy sets; and B- is the
output fuzzy set, and y is the output variable. Having multiple
behaviors, which are all running concurrently, leads to
situations where several command outputs may be produced
simultaneously. Therefore, the main advantage of using Fuzzy
Logic for navigation is that it allows for the easy combination of
various behaviors through a command fusion process instead of
using fixed parameters in the entire process.
The design of a Fuzzy Logic controller starts with the definition
of the membership functions for the output variable, here, SL.
Currently, seven empirically determined membership functions
are used for SL in our prototype, as shown in Figure 3. The
fuzzy language for this fuzzy set is divided into a range of
quantities such as: Zero, Very Short, Short, Normal, Semi-Long,
Long, and Very Long; vertical axis in Figure 3 indicates the
degree of membership of SL in the corresponding fuzzy set
(fist.)-
Defining the shape, the membership functions, and the bounds
of these quantities is a design problem, but the attributes of the
system will not be changed significantly if the membership
functions are modified slightly. The value of the membership
function indicates the degree of membership of SL to the fuzzy
set. If the membership value is 1 for one of the fuzzy sets, the
SL is perfectly representative of the set, and if it is 0, the
quantity is not at all a member of the set. Any value between 1
and 0 indicates a partial membership. A better way to make SL
a fuzzy set is to allow the membership functions to change
gradually from one quantity to the next one. Then, the real
power of the Fuzzy logic comes from the ability to integrate
these partial membership values in a way that permits a good
balance between membership functions.
For reliable SL/SD results, the KBS system must be sufficiently
trained, meaning that sufficient amount of calibration data must
be either stored in the memory or provided during the actual
navigation task, before the GPS signals are blocked. For the
ANN module training, different terrains slopes/configuration
and types of surfaces must be included, for a representative
number of operators, to derive a reliable predictive model;
obviously, if the system is calibrated under circumstance totally
different from the actual navigation task, the results will be
much worse than the examples provided here. Similarly, the FL
modules requires a large sample of representative data where
various human dynamics types are included in various
environmental conditions and terrain configurations, to derive
the appropriate fuzzy rules for the membership functions that
will be used to predict the model parameters once GPS signals
are blocked. The additional benefit of FL is that the actual
behavior of the mobile operator can be predicted, that is, if the
person is running, walking, stumbling, climbing, etc., and that
might be useful information in particular in combat or
emergency situation, and can be wirelessly transmitted to an
operational center (not implemented in our prototype).
An additional use of FL in our implementation is the adaptive
Extended Kalman Filter where the adaptivity scheme is based
on Fuzzy Logic rules (see, e.g., Sasiadek et al., 2000;
Moafipoor, 2008). In this approach, the pseudorange practical
1 m j
covariance, C, = — Y. e, e , and the actual covariance
k mi=1 k k
— T
(covariance of innovation) from the EKF, s k = H k P k H^ ,
are compared, and the level of the difference between them is
tested using fuzzy rules to decide if the measurement covariance
matrix R k should be modified (adapted to the current state of
system sensors). H k is the observation design matrix, P^ is the
predicted covariance, and e k is the innovation vector. The
system calibration mode with the KBS module is illustrated in
Figure 4.
Figure 3. SL membership function.