Recognizing affect based on body movement and contextual information
Recognizing emotions through bodily expression is a relatively emerging field. There is yet no consensus or standard for an agreed upon methodology or framework yet. As there is little work on non-acted expressions for bodily expression, there is further less for ambient intelligence. However there...
Saved in:
| Main Author: | |
|---|---|
| Format: | text |
| Language: | English |
| Published: |
Animo Repository
2016
|
| Online Access: | https://animorepository.dlsu.edu.ph/etd_masteral/5211 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | De La Salle University |
| Language: | English |
| Summary: | Recognizing emotions through bodily expression is a relatively emerging field. There is yet no consensus or standard for an agreed upon methodology or framework yet. As there is little work on non-acted expressions for bodily expression, there is further less for ambient intelligence. However there is also an emergence in overlaps between methods and practices and there are e orts in developing corpora for the analysis of spontaneous affect using bodily expression. While recognition performance reached up to 95% in related studies, most of research conducted with the body as a modality used acted data. Research with natural data reach above chance agreement with observers and models, but do not perform as well as with acted data. The current research looks into context as a possible source of information that can be used to improve the classification. However, current methods do not take largely context into consideration.
A representation was defined for context and is annotated manually. An active serves as the emotion stimulus and contextual information. Context is represented as discrete categories describing events and task related states. Posture and Movement features are extracted from Kinect skeleton data. Models were trained to classify Valence, Arousal and Activity Intensity from posture, movement and context. Various model configurations were analyzed to investigate how context and bodily expression affect the classification of emotions. Various relationships were observed from the models. On average, the performance falls between 50% to 60% for train and test sets. However, given the right configuration, models trained to classify Valence, Arousal and Activity Intensity, reached up to 70% to 80% in terms of accuracy. |
|---|