Year: 2024
Type: dissertation
Abstract: The study of mother-infant interactions is fundamental to understanding early childhood development, as these interactions influence emotional, social, and cognitive development.Motherinfant interactions help form secure attachment bonds, advance language skills, and nurture social competencies.Traditionally, the analysis of these interactions has relied on video observations and manual coding by trained psychologists, a method that is labor-intensive and costly.Therefore, there is an urgent need for cost-efficient and dynamic approaches.This study introduces an advanced approach using Machine Learning (ML) and Deep Learning (DL) technologies to automate the analysis of mother-infant interactions, with a focus on video action recognition.We employ a novel hybrid architecture integrating Pose-Guided Motion History Analysis (PGMHA) and Transformer-Based Sequence Modeling (TBSM) to enhance action recognition in mother-infant interaction videos.A transformer-based model is used to predict tactile stimuli, Voice Activity Detection (VAD) for auditory stimuli, and Computer Vision techniques to analyze visual stimuli, focusing on gaze direction detection.The proposed model achieves 74.8% accuracy in identifying and categorizing maternal behaviors, demonstrating its efficacy in recognizing interaction patterns.This approach provides new possibilities for psychologists to efficiently analyze large volumes of mother-infant interaction data with sufficient accuracy, reducing the time and resources needed for manual coding.By automating the recognition of key interaction behaviors, it supports more precise and scalable research, potentially leading to earlier identification of parenting and dyadic interaction problems and more targeted interventions.
Author: Mohammad, Fareeda Madeen Gouher
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7133002067
