Using Computer Vision to Annotate Video-Recoded Direct Observation of Physical Behavior.

Area of Science:

• Computer Vision
• Machine Learning
• Human Movement Analysis

Background:

• Direct observation is a gold standard for physical behavior assessment but is labor-intensive and costly.
• Automated methods are needed to improve the efficiency and scalability of physical behavior analysis.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for recognizing physical behavior and location from video data.
• To assess the accuracy of ML models across different physical behavior taxonomies.

Main Methods:

• Adults (N=26) were video-recorded in their natural environment.
• Videos were annotated for sedentary status, activity type, and intensity using established taxonomies.
• Four ML approaches were trained and validated on 90% of video data, with final accuracy tested on the remaining 10%.

Main Results:

• ML models achieved 87.4% accuracy for sedentary vs. non-sedentary classification.
• Accuracy for activity type and intensity classification was 63.1% and 68.6%, respectively.
• Computer vision effectively automates physical behavior annotation, reducing time and labor.

Conclusions:

• Machine learning models demonstrate feasibility for annotating physical behavior from videos.
• This approach offers a scalable alternative to direct observation for human movement analysis.
• Future work should involve larger datasets and video fragment analysis for enhanced model performance.