Emotion and motion: Toward emotion recognition based on standing and walking

Area of Science:

• Human-computer interaction
• Affective computing
• Biomedical engineering

Background:

• Emotion recognition is crucial for effective human-computer interaction.
• Previous studies often used acted emotions and limited motion parameters, hindering real-world applicability.
• Naturalistic body motion offers a richer, more authentic source for emotion recognition.

Purpose of the Study:

• To develop and evaluate an approach for emotion recognition using naturalistic body motion.
• To investigate the effectiveness of various machine learning models in classifying emotions from a comprehensive set of motion parameters.
• To advance emotion recognition capabilities for practical human-machine interaction.

Main Methods:

• A laboratory experiment with 24 participants exposed to five emotional states (happiness, relaxation, fear, sadness, neutral) induced by movies.
• Motion capture system and force plate used to collect posture, motion, and center of pressure data during standing and walking.
• Six machine learning models (k-NN, decision tree, logistic regression, SVM variants) trained and evaluated using 229 motion parameters.

Main Results:

• Traditional statistical analysis revealed limited significant effects of emotion on individual motion parameters (7 out of 229).
• A decision tree model utilizing 25 parameters achieved the highest average accuracy of 45.8%, significantly outperforming random chance.
• This accuracy surpasses previous studies, attributed to the naturalistic setting and comprehensive parameter set.

Conclusions:

• Emotion recognition from natural body motion is feasible using machine learning, despite limited effects of individual parameters.
• Machine learning models, especially decision trees, are valuable tools for advancing emotion recognition in realistic scenarios.
• This research provides a foundation for developing practical emotion recognition devices for human-machine interaction.