Skeleton-Based Infant Activity Recognition: A Study of Sequence, Graph, and Time-Frequency Representations

Accurately recognizing developmental milestones in the first year of life is essential for monitoring infant growth and supporting early cognitive and motor development. While Human Activity Recognition (HAR) has been extensively studied in adults, infant-specific HAR remains underexplored due to the distinct movement patterns of infants, higher data variability, and annotation difficulties. In this work, we present a comparative study on skeleton-based infant activity recognition using three distinct data representation strategies: sequence-based, graph-based, and time-frequency (scalogram-based) approaches. We design and evaluate ten models: five for sequence representations, one graph neural network, and four convolutional models for time-frequency data, each aimed at capturing the unique dynamics of infant motion. Skeleton data were extracted from RGB video through a pose estimation pipeline, enabling non-invasive and scalable analysis. Experimental results show that sequence-based models outperform graph and scalogram-based counterparts, highlighting their effectiveness in capturing temporal dynamics essential for accurate recognition of infant activities. This study provides a structured benchmark comparing data representation techniques in infant HAR and offers practical insights for designing early monitoring tools to support child development assessment.