Baby and Childcare Context Recognition

Overview

Recording and sharing childcare information is crucial for accurately assessing a child's health status, yet manual recording presents a significant burden for parents. This project develops methods to automatically detect and recognize baby and childcare activities using off-the-shelf wearable devices and inertial sensors.

Approach

• Infant activity recognition (IEEE Pervasive Computing 2024): Multi-label classification model using a chest-mounted low-sampling-rate accelerometer to recognize 14 daily activities of infants aged 6-24 months, with 25 time- and frequency-domain features
• Preliminary investigation (UbiComp 2023): Machine learning model classifying 8 typical daily behaviors (sleeping, crawling, walking, standing, sitting, drinking milk, eating, being held) from chest-mounted accelerometer data
• Childcare activity detection (SmartComp 2022): Detecting 9 parenting activities using motion-sensor data from an off-the-shelf smartwatch worn by parents

Results

• Infant activity recognition: Achieved over 88% F1 score in the best case for multi-label classification of 14 activities, validated with data from 24 infants
• Preliminary model: Nearly 80% accuracy for 8-class classification with data from 10 infants over ~18 hours
• Parenting activity detection: 71% accuracy (F1: 0.66) for 9 childcare activities in laboratory evaluation

Significance

This research demonstrates the feasibility of automatic, non-intrusive recording of infant and childcare activities using a single commercially available sensor. By significantly reducing the manual recording burden for parents, it enables continuous health monitoring and timely response in case of illness or emergencies.