Nan Bu, Masaru Okamoto, Toshito Tsuji
IEEE Transactions on Robotics,
Vol.25, No.3, June 2009
Summary
The paper discusses the usage of
electromyogram (EMG) signals, classified simultaneously by probabilistic neural
networks, for human motion prediction in a human-robot cooperation environment.
EMG signals are electrical manifestations of the activity muscles measured at
the skin surface and in the literature difference application are described are
done, significant is Tsuju et. Al, who introduce an Entropy based decision rule
to reduce misclassifications, the system imposes that when entropy exceeds a certain
value, then the motion decision rule must suspend the judgment. The paper
proposes a Bayesian Network for motion prediction, it presents the following
advantages: 1) Being a probabilistic expression, it is good in dealing with
uncertainty, which is typical of human behavior decision; 2) The network
structure can be created from data or devised by hand; 3) the conditional
probability tables, which compose the parameters of the problem, can be obtained
from the data itself. Unlike the usage of Petri Nets (Fukuda et. al), the usage
of Bayesian Networks doesn’t need predefined detailed sequences, but
conditional probabilities extracted from the data.
·
Model
The Bayesian Network is a
graphical notation that encodes conditional dependence relationships among a set
of events, it is dependent on nodes variables (V), arcs between nodes (A)
and probabilities associated with each node. The model considers mc, motion at
current step, mp, motion for previous step, lp, location
of motion mp, hp, user’s hand position at previous step,
P(mp), probability of motion, P(hp), probability of hand
position at previous step, P(lp|hp), conditional
probability of location of location at the previous step given the previous
hand position, P(mc|mp,lp), conditional
probability of the current motion with respect to motion and location
information at previous step, this can be calculated by knowing the number of
samples in a database as shown.
P(lp|hp)
is a continuous probability distribution (due to the presence of hp,
which is a continuous variable) and can be calculated through Bayes’ Law.
EMG signals are collected
from a certain number of electrodes, they are rectifies, filtered by a second
order Butterworth filter and digitized with a sampling frequency fs,
in the end a feature vector x(t) is
obtained and used as pattern to be classified by a probabilistic Neural Network
(LLGMN – Log-linearized Gaussian mixture network), giving O(t), the posterior
probability vector, as a final output. The amplitude of the EMG indicated the
force information FEMG(t). The motion belief appears to be very
important, it initializes the model with a value of 1/M, with M indicating the
sample size, and later calculated as for P(mc|mp,lp).
Motion decision is taking onlywhen the force FEMG(t) is more or
equal than Fth, which is an indicative value for this purpose. Also
the Entropy may be calculated and when is surpasses a certain fixed value Hd,
the motions will be suspended. During motion a probability of motion is used to
and is dependent on motion, output and a weight wm. The importance
of wm is relevant to the fact that it introduces α as value which determines the influence in motion
preditiction combination (the larger and the greater is the influence).
Key Concepts
EEG signals for robotic
control, Bayesian Network for motion prediction
Key Results
An experiment on the system
has proved the improvement of integrating BN in LLGMN, obtaining a
statistically relevance reduction in misclassification. The model appears still
to be improved, due mainly to the difficulty in human behavior prediction, but
it definitely can provide, through BN, more robustness and reliabil
No comments:
Post a Comment