Information Technology Reference
In-Depth Information
Fig. 7.2 Online HAR algorithm stages. This illustration depict schematically the input and output
of each block
The first stage, signal conditioning and feature extraction , comprises data
acquisition and signal conditioning from the inertial sensors to obtain the features
that characterize each activity sample. These features become the input of second
stage, multiclass SVM with probability estimates , where they are evaluated for activ-
ity prediction. From each sample, an array is extracted to indicate the probability
of belonging to the learned classes. In the third stage, temporal activity filtering ,
these probabilities are then joined with the predictions of previous activity samples
and processed by means of temporal filtering. This is achieved by applying a set of
defined heuristic filters that only allow natural sequences of activities. They consider
PTs and also unknown conditions (e.g. when output probabilities are marginal) for
this purpose.
Both recognitionmethods (PTA-6Aand PTA-7A) fit in the same pipeline changing
only in the number of trained classes of the MC-L1-SVM.
7.3.1 Signal Conditioning and Feature Extraction
The inputs of the PTA-HAR system are the raw triaxial linear acceleration a r (
t
)
and
angular velocity ω r (
time signals. These are read at a constant frequency of 50Hz in
the process which performs the signal conditioning: the Process I nertial Si
t
)
()
function in Algorithm3. The set of filters used are for this task are: (i) noise reduction,
whose transfer function is represented by H 1 ()
g
nals
, is achieved by applying a third-order
median filter and a third-order low-pass Butterworth filter (cutoff frequency = 20Hz).
These filters allow to obtain clean triaxial acceleration a
˄ (
t
)
signal. Angular velocity
ω (
is additionally processed with a high pass filter (0.3Hz cutoff frequency),
represented by the transfer function H 2 ()
t
)
, in order to remove any bias in the signal.
(ii) The segmentation of the acceleration signal into gravity g (
t
)
and acceleration
due to body motion a
(
t
)
. This is possible by also high-pass filtering the acceleration
a ˄ (
t
)
with H 2 ()
to obtain a
(
t
)
. g (
t
)
is subsequently found by subtracting a
(
t
)
from
a ˄ (
(Refer to Sect. 4.3.2 for more details).
In addition, the Online Prediction
t
)
function, which is in charge of the recog-
nition of activities, is periodically executed to obtain and classify window samples
( A
()
,
,
) , g (
) , ω (
) over
a period T . Its periodicity satisfies the sliding-windows criteria: a time span of 2.56 s
G
) extracted from the filtered triaxial inertial signals ( a
(
t
t
t
)
 
Search WWH ::




Custom Search