Beyond proximity: A computer vision (CV) framework for quantifying occupant interactions and airborne exposure in indoor environments
Building and Environment · 2026-07 · 299 · 114696
DOI: 10.1016/j.buildenv.2026.114696 ↗Kurz gefasst
Fachartikel in der Zeitschrift Building and Environment, 2026-07. Verfasst von Gautam Vanama und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 48 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0). Gefördert von: Johns Hopkins University, NIH.
Abstract (Originalsprache)
Indoor airborne exposure assessment depends on both ventilation conditions and occupant interactions, including proximity, duration, face-to-face alignment, and respiratory activity. However, most monitoring approaches reduce activity to occupancy counts or distance thresholds that treat all within-range encounters as equivalent. This study presents a multi-factor framework combining computer vision (CV)-derived interaction geometry, including mouth-to-mouth distance, and face-to-face orientation, with respiratory activity characterization for interaction-aware exposure stratification. Deployed on twenty source-receptor pairs from two controlled cohorts, the framework uses multi-camera triangulation for 3D mouth position estimation, multi-view head orientation inference for face-to-face alignment, and privacy-preserving re-identification for cross-view association. These interaction inputs parameterize a Wells-Riley-inspired relative aerosol exposure proxy intended for exposure stratification rather than infection prediction. Compared with distance-only assessment, the full multi-factor model increased across-pair discrimination (geometric P90/P10) from 42.5 (95% CI: 15.5–107.4) to 2,006 (95% CI: 471–39,068), corresponding to a roughly 50 × improvement observed in 99.3% of bootstrap replicates. Respiratory activity-dependent source variation provided the largest contrast, while face-to-face orientation further distinguished pairs at similar distances and duration by directional alignment and posture. The CV pipeline achieved 96% of frames within ±2 cm mouth-to-mouth distance error and re-identification IDF1 up to 98%. These results show that jointly measuring interaction geometry and respiratory activity captures exposure-relevant variation missed by proximity-only methods, providing a foundation for interaction-aware indoor exposure assessment and integration with ventilation strategies. Under a conservative source-rate parameterization, the multi-factor advantage was 5.0 ×, supporting the conclusion that the improvement is structurally robust to emission-rate assumptions.
Förderung
- Johns Hopkins University
- NIH
Zitieren
APA
Vanama, G., Zhu, S., Milton, D. K., & Srebric, J. (2026). Beyond proximity: A computer vision (CV) framework for quantifying occupant interactions and airborne exposure in indoor environments. Building and Environment, 299, 114696. https://doi.org/10.1016/j.buildenv.2026.114696
Vancouver
Vanama G, Zhu S, Milton DK, Srebric J. Beyond proximity: A computer vision (CV) framework for quantifying occupant interactions and airborne exposure in indoor environments. Building and Environment. 2026;299:114696. doi:10.1016/j.buildenv.2026.114696
BibTeX
@article{Vanama2026,
author = {Vanama, Gautam and Zhu, Shengwei and Milton, Donald K. and Srebric, Jelena},
title = {Beyond proximity: A computer vision (CV) framework for quantifying occupant interactions and airborne exposure in indoor environments},
journal = {Building and Environment},
year = {2026},
volume = {299},
pages = {114696},
doi = {10.1016/j.buildenv.2026.114696}
}
EUVIMED