Automatic Recognition of Non-Acted Affective Postures.

Area of Science:

• Human-computer interaction
• Affective computing
• Computer vision

Background:

• Understanding and recognizing human emotions (affective states) is crucial for effective human-technology interaction.
• Current methods often rely on acted emotions, which may not reflect real-world affective expressions.
• There is a need for technology that can accurately perceive users' affective states in naturalistic contexts.

Purpose of the Study:

• To develop and validate automatic recognition models for affective states and dimensions using non-acted body postures.
• To assess human observer agreement on affective labels derived from avatar postures.
• To evaluate the generalizability of automatic recognition models to new observers and postures.

Main Methods:

• Trained and tested automatic recognition models on non-acted body postures from a body-movement-based video game.
• Used faceless avatars to represent postures and collected human observer agreement on affective labels.
• Employed random repeated subsampling validation to test model generalization using observer-assigned labels as ground truth.

Main Results:

• Automatic recognition models successfully identified affective states and dimensions from non-acted body postures.
• Human observers showed agreement above chance level when labeling affective states from avatar postures.
• The developed models achieved recognition accuracy comparable to human observer agreement rates.

Conclusions:

• Automatic recognition of affective states from non-acted body postures is feasible and effective.
• Technology can be endowed with the ability to recognize users' affective states for improved interaction.
• This research advances affective computing by demonstrating robust emotion recognition from naturalistic body movements.