Graphics Reference
In-Depth Information
deviations are formed for signals, and signal increases or decreases
calculated with these. To do so, statistical methods such as
t
-tests and
variance analysis are used and effect strengths calculated (Kreibig
et al., 2007, 2010; Stemmler and Wacker, 2010; Schupp et al., 2004;
Bradley, 2009). In the affective computing area, generally spoken, the
raw signals are initially subjected to a (1)
pre-processing
and then a
(2)
feature extraction
(f
#
). The number of extracted features ranges
from 13 (Haag et al., 2004) to 110 (Kim and André, 2008).
There is no consensus yet about which feature extraction is better.
As an example, the extraction of Gu et al. (2008) will be mapped. Gu
et al. used the following formulas for corrugator and Zygomaticus
EMG, BVP, SCL, temperature (TMP) and electrocardiogram parameters
(ECG) for each of the six parameters and extracted 36 features:
The mean of
x
(
n
)
The standard derivation of
x
(
n
)
The mean of the absolute values of the first differences of
x
(
n
)
The mean of the absolute values of the first differences of
normalized
x
(
n
)
The mean of the absolute values of the second differences of
x
(
n
)
The mean of the absolute values of the second differences of the
normalized
x
(
n
)
What follows is an automatic (3)
feature selection
. Gu et al. (under
review) were able to show with the Sequential Floating Forward
Search (SFFS) algorithm that, with regard to valence, the accuracy
and robustness reaches their highest levels at 10 features and starts
to decrease at 20 features, but arousal already reaches its highest level
at 5 features and starts to decrease at 22 features. The problem with
the feature selection is that these features are individual-specific and
trans-situational dependent. Nevertheless, Kolodyazhniy et al. (2011)
selected features that are inter-individually and trans-situationally
robust: SCL, Corrugator-EMG, Zygomathicus-EMG, pCO
2
(end-tidal
carbon dioxide partial pressure) and PEP (pre-ejection period).
The last step is the (4)
classification
(LDP, SVM, MLP, etc.)
or hybrid classification, respectively. Overall, however, based on
current findings, it can be said that psychobiological signals have the
advantage that they can be obtainable in a permanent fashion and
regardless of the location. It is absolutely necessary, however, that
sensors are “comfortable”. The psychobiological gathering of data
tends to require individual-specific processing (Walter et al., 2013a,
Böck et al., 2012). A trans-situational, robust feature selection still
currently presents a problem (Walter et al., 2013a).
Search WWH ::
Custom Search