Unsupervised End-to-End Deep Model for Newborn and Infant Activity Recognition

Area of Science:

• Biomedical Engineering
• Machine Learning
• Infant Care Technology

Background:

• Existing human activity recognition (HAR) research primarily targets adults, overlooking the unique needs of infants.
• Infants' inability to communicate verbally makes HAR crucial for their safety and well-being.
• Infant activities differ significantly from adult activities in type and intensity, necessitating specialized study.

Purpose of the Study:

• To develop a novel method for recognizing newborn and infant activities using sensor data.
• To classify four distinct infant activities: sleeping, distressed movement, normal movement, and externally induced movement.
• To address the gap in HAR research concerning the specific behaviors of infants.

Main Methods:

• Collected 11 hours of video and synchronized accelerometer data from 10 infant subjects.
• Proposed an end-to-end deep learning model integrating an autoencoder and k-means clustering.
• Employed an unsupervised learning approach for model training.

Main Results:

• The proposed model achieved a balanced accuracy of 0.96.
• The model demonstrated a strong performance with an F-1 score of 0.95.
• The system effectively distinguished between various infant activities based on sensor data.

Conclusions:

• The developed deep learning model shows high efficacy in infant activity recognition.
• This technology can significantly contribute to infant safety and health monitoring systems.
• Unsupervised learning with autoencoders and k-means clustering is a viable approach for infant behavior analysis.