Estimating vehicular emissions by applying deep learning on video camera scenes

Abstract

Despite advances in electrification, urban traffic remains a substantial source of air pollution and poses a major threat to public health. Addressing the spatial and temporal heterogeneity of vehicular emissions requires scalable methodologies. Here, we present a deep learning-based framework that analyzes street video footage to derive vehicle-resolved emission estimates for high-resolution urban monitoring. Our approach first classifies vehicles using a dataset of 2.3 million images, covering 4,923 car models. We then track vehicles and extract their speed and acceleration, which are used to compute emissions via a modified version of the COPERT model. Applying this method in Amsterdam and Abu Dhabi, we demonstrate its effectiveness for real-time traffic flow analysis and its potential for scalability. Our results provide a standardized approach to high-resolution urban emission monitoring, with implications for data-driven urban policy and environmental regulation worldwide. Ashutosh Kumar and researchers use vehicle-mounted cameras and deep learning to estimate vehicle-level emissions. Tests in Amsterdam and Abu Dhabi show how the method can support high-resolution urban monitoring.