Deep Learning Approaches for Infant Activity Recognition: a Comparative Study of Sequence Models and Ensemble Learning Techniques

Abstract

Supporting healthy growth and cognitive development in infants necessitates accurate recognition and monitoring of developmental milestones during the first year of life. Effective monitoring plays a pivotal role in this process, enabling parents and caregivers to provide personalized support and promote overall well-being. Although Human Activity Recognition (HAR) has been extensively studied, the majority of research has primarily targeted adult populations, leaving infant-specific HAR under-explored. This research advances infant activity recognition through three major contributions: (1) enhancing an existing dataset under the supervision of a pediatric physiotherapist to ensure clinical relevance and accuracy, (2) conducting a comprehensive evaluation of different sequence deep learning models, and (3) investigating the efficacy of ensemble learning through four distinct strategies. Experimental results demonstrate that ensemble learning outperforms individual models, achieving a 4.87% improvement in recall over the best-performing baseline, CNN-LSTM, and a 5.31% improvement over the most stable model, Bi-LSTM. These findings offer promising implications for the design of intelligent infant monitoring systems aimed at early detection of developmental milestones and improved longitudinal assessment.