Estimating vehicular emissions by applying deep learning on video camera scenes

Urban traffic remains an important source of anthropogenic carbon emissions, contributing significantly to climate change. Traffic-related air pollution poses a great public health threat to the urban population. However, cities globally have yet to devise a universal regulation approach due to traffic emissions' spatio-temporal heterogeneity. This study introduces an approach employing computer vision to map vehicle pollutants at the vehicle level. Our methodology utilizes a unique car model classification dataset, comprising 2.2 million images, enabling the classification of 4,923 car models. In addition, we associate each vehicle with its emissions using the EU standard vehicular emission model. To improve emission estimation accuracy, we use a modified version of COPERT that incorporates both velocity and acceleration. This enhancement provides a more refined representation of real-world driving conditions. By leveraging vehicle tracking and distance measurement from video footage, we estimate both speed and acceleration for all vehicles in the scene, offering a more comprehensive understanding of traffic dynamics and their influence on emissions. The application of our method in Amsterdam demonstrates the potential for real-time traffic flow detection and emission estimation worldwide. Our findings contribute to addressing the imperative vehicular emissions measurement and regulation challenges in dense urban areas globally.