Development and evaluation of a computer vision algorithm for quantification of children's microactivities.

Area of Science:

• Environmental health
• Computer vision
• Child development

Background:

• Estimating children's microactivity frequencies (e.g., hand-to-mouth contact) is crucial for environmental exposure modeling.
• Manual video analysis for these behaviors is time-consuming and costly.

Purpose of the Study:

• To develop and evaluate a computer vision model for quantifying microactivities in young children.
• To automate the measurement of hand-to-mouth and object-to-mouth contact events.

Main Methods:

• Trained a computer vision model using video footage of 25 children (6-18 months) in home environments.
• Utilized 3D key point tracking to assess child pose and measure hand-to-mouth distance.
• Defined "contact events" when hand-to-mouth distance met a minimum threshold, comparing model outputs to human coding.

Main Results:

• The computer vision model accurately recognized children's microactivities.
• Detected contact events with 96-99% accuracy on a second-level basis, showing minimal counting errors.
• Observed significantly higher rates of object-to-mouth contacts (27 contacts/h) compared to hand-to-mouth contacts (3 contacts/h).

Conclusions:

• A computer vision method was successfully developed and validated for quantifying children's microactivities.
• This automated approach significantly reduces the cost and effort required for generating microactivity data.
• The findings support improved soil and dust exposure modeling for children.