Recognizing Construction Vehicle Activity in Low-Frequency Surveillance Imagery and Its Association with Local Air Quality

Construction sites are major urban emitters, yet dense, high-quality pollution sensing remains costly. This paper develops a three-stage machine learning pipeline that uses sparse on-site surveillance imagery to detect and classify construction vehicle activity and quantify its association to ambient air quality. The pipeline comprises (i) a YOLOv9-based detection model, (ii) an activity classifier evaluated across three contrasting architectures, Vision Transformer (ViT), ResNet-50, and SVM, and (iii) a linear regression relating detected activities to local air quality indicators. Despite operating on low-frequency (5-minute) imagery under real-world conditions, the models achieve stateof-the-art performance in both detection and activity inference. Regression analyses reveal statistically significant but limited correlation between vehicle activity and local pollutant concentrations, suggesting dominant external sources or unobserved confounders. These findings illustrate the feasibility of passive, vision-based environmental sensing in constrained urban deployments and motivate tighter integration of ubiquitous computing with sustainable construction monitoring.