AI System In The Real World: Capture Children Food Exposure Using Wearable Cameras

Abstract

Children’s dietary habits are influenced by complex factors. Identifying community-level influencers and measuring their effect is traditionally based on self-reported data and prone to measurement error. In addition, researchers are usually unable to accurately capture children's dietary habits and food exposure using traditional surveying techniques. We propose to develop a culturally acceptable AI based data collection system that objectively captures school-children’s exposure to food items, food ads, markets, etc... We engaged students, parents and school staff in a user-centered study for the food exposure design of school children through a feasible model in Beirut and Tunis. Findings suggest that wearable cameras are suitable when used for a limited period (24 hours max). Nonetheless, some ethical challenges were raised related to privacy, confidentiality and anonymity along with suggestions and solutions on how to address them with technology in a scientific objective manner and how to make such an AI system acceptable so that it can be used in similar research studies. We also survey a list of the most popular wearable cameras and present their pros and cons. We train a machine learning model for automatically detecting food-exposure images and blurring faces that will be automatically captured by wearable cameras. This report also discusses how we automatically collected from the Web the training dataset for training such an ML model and how we managed to overcome any possible bias in our model and report on its performance. Furthermore, we explain and expose the software tool that packs the AI system that we developed and we showcase its simple-to-use interface and we report on its throughput, efficiency and hardware requirement. Finally, we discuss how we deployed this AI-based system in the real world in a real study and we provide a high level of analysis on the collected data from the real world.