Computer vision for extraction of environmental characteristics from street images: a scoping review of methods and applications
Journal of transport & health 2025
Kokka K, Huang Y, Itova I, Schönlieb CB, Foley L, Woodcock J, Burgoine T
DOI : 10.1016/j.jth.2025.102209
URL : https://doi.org/10.1016/j.jth.2025.102209
Abstract
Introduction
The built environment can influence diet, participation in physical activity, related non-communicable disease outcomes and mortality. Therefore, characteristics of the built environment have been the subject of much international public health research. However, in many contexts the ability to measure environmental exposure is limited by data availability. In-person street audits can provide detail but are costly at scale, while virtual, desk-based audits are also resource-intensive. Computer vision (CV), powered by deep learning, which can automatically extract data on environmental characteristics from street images, is a potentially powerful alternative. In this systematic scoping review, we explored the uses, models and performance of CV for the detection of environmental characteristics of potential relevance to diet and physical activity behaviours from street images.
Methods
Following an adapted version of Arksey and O'Malley's review process, we used eight diverse databases to identify 11,221 records. Eligible studies were published 2020–2023, reported in English and focused on CV models to identify objects relevant to diet or physical activity from street images. After title, abstract, and full-text screening, we included 106 studies in our review. We conducted a narrative synthesis of findings, supported by harvest plots.
Results
Most studies employed pre-trained, segmentation models such as DeepLabv3 and YOLO, based on Cityscapes and MS COCO benchmark datasets. Applications of CV have been concentrated in the United States and China, and in high income countries more generally. CV was used to detect data on 40 broader environmental characteristics relating to the built, natural, transport and food environments. Less than half of the studies we found reported model accuracy.
Conclusions
Our findings indicate the potential of CV in public health research. However, it is a concern that few studies have reported CV model performance. We provide some minimum reporting recommendations concerned with the use of CV to maintain trust and transparency in public health research.