Modeling Full-Body Movements for Emotion Classification

This research article investigates the classification of emotions based on full-body movements using a novel Convolutional Neural Network (CNN) architecture. The model incorporates two shallow networks that process 8-bit RGB images derived from 3D positional data captured over different time intervals. One network focuses on coarse-grained modeling while the other emphasizes fine-grained modeling, leading to improved classification results for a dataset featuring professional dancers expressing four emotions: anger, happiness, sadness, and insecurity. The study also examines the impact of data chunk duration, overlapping, image size, and various data augmentation strategies on classification performance. Results indicate that longer data chunk durations enhance recognition accuracy, particularly when balanced data augmentation techniques are applied. The proposed method was further validated against other motion capture datasets, demonstrating superior performance compared to existing approaches. The findings suggest that the method generalizes well across diverse contexts and settings, contributing to the field of emotion recognition from body movements.