Automated High-Frequency Observations of Physical Activity Using Computer Vision | JoVE Visualize

Automated High-Frequency Observations of Physical Activity Using Computer Vision

Jordan A Carlson, B O Liu¹, James F Sallis²

¹Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA.

Medicine and Science in Sports and Exercise|March 17, 2020

View abstract on PubMed

Summary

Computer vision algorithms accurately assess physical activity in public spaces. The Ecological Video Identification of Physical Activity (EVIP) system shows promise for automated, efficient, and reliable activity monitoring.

Area of Science:

* Kinesiology and Public Health
* Computer Vision and Machine Learning
* Behavioral Science and Epidemiology

Background:

* Accurate assessment of physical activity in real-world settings is crucial for public health research and interventions.
* Traditional methods like direct observation and accelerometry can be resource-intensive and may not capture the full ecological context.

Purpose of the Study:

* To validate the Ecological Video Identification of Physical Activity (EVIP) computer vision algorithms.
* To assess the automated, video-based ecological measurement of physical activity in parks and schoolyards.

Main Methods:

* Trained and tested computer vision algorithms using 27 hours of video data from stationary overhead cameras.
* Utilized accelerometry data as ground truth for classifying physical activity levels (moderate-to-vigorous vs. sedentary/light).
* Compared EVIP performance against ground truth and traditional System for Observing Play and Recreation in Communities (SOPARC) observations.

Main Results:

* EVIP demonstrated high agreement with ground truth for counting people (CCC=0.88) and moderate agreement for identifying active individuals (CCC=0.55).
* EVIP outperformed SOPARC observations in accuracy for both people counts and activity levels.
* Algorithm error was not significantly correlated with environmental factors like camera placement or lighting conditions.

Conclusions:

* Computer vision algorithms offer a promising automated approach for assessing setting-based physical activity.
* EVIP technology has the potential to reduce manpower needs compared to human observation.
* Automated systems can provide continuous monitoring and data for informing public health interventions.