Graphics Reference
In-Depth Information
and Rösner, 2010). Emotion-related annotated data is classified
through a dynamic recognition process based on Hiden Markov
models, neuronal networks, and kernel-based processes as well as
support vector machines (SVM) and kernel logistic regression (KLR).
The language-independent classification of basic emotions lies at
approximately 75%, and the language-dependent classification at 93%,
which is very precise, with regard to artificially created language
material with emotional prosody (Wagner, 2005), if the language
signals come from a defined interaction and are recorded without any
acoustic interference.
6.6 Psychomotor functions: Gestures, body movements
and attention focus
The automatic recognition of gestures generally takes place in three
steps: (1) the object detection, (2) the chronologically recursive
filtering (tracking) and (3) the classification and/or verification (Bar-
Shalom and Li, 1995). (1) In the detection step, the measured data is
individualized, which means that object hypotheses with individually
measured data are extracted from the measured data, which then
must be allocated to the known objects. (2) The chronological filtering
then, in the second step, uses a multi-instance filter approach, in
which every object is individually assigned a dynamic filter. For
chronological filtering processes, approximations of the generally
recursive Bayes estimation are used (Bar-Shalom and Li, 1995). (3) In
the area of gesture recognition, numerous classification methods have
been developed ((Morguest, 2000), (Corradini and Gross, 2000) and
(Barth and Herpers, 2005)). Once the static and dynamic gestures have
been recognized, they can be classified. For static gestures, model-
based processes (Stenger et al., 2001), Active Shape Models (Wimmer
and Radig, 2007), or feature-based processes in connection with a
classification algorithm (e.g. Bayesian Networks (Ong and Ranganath,
2003) or artificial neuronal network, 2001) can be used. In the dynamic
methods, the focus is on the analysis of a chronological sequence of
individual images to determine the movement trajectory. To do so,
image and pattern recognition methods for time-dependent features
must be used in order to analyze the information from the video
sequence. The typical sequential process flow for gesture classification
includes segmentation, feature recognition and feature extraction.
Search WWH ::
Custom Search