The gas pollution of the air by vehicle emissions in a sustainable, smart city is an undeniable and urgent problem that requires new methods of solution. In order to take administrative measures to improve air quality, it is necessary to have a reliable tool for instant assessment of current air pollution at road intersections as places of the largest accumulation of vehicles. This article describes a mathematical model and its software implementation based on neural network technology, which allows continuous monitoring of emissions of nine types of pollutants from various categories of standing and moving vehicles with such parameters as speed, coordinates, and idle time. Authors developed a dataset for the dynamic neural network training, which consists of 60,000 labeled images. The model calculates the pollution level of an air basin in an area determined by the visibility zone of an outdoor video surveillance camera and a height of 2 meters. Unlike existing models, the proposed solution works in real time mode, can be embedded into the existing infrastructure for monitoring road intersections, and takes into account current weather conditions: wind strength and direction. This allowed the authors to verify the results of calculations with instrumental measurements of a mobile environmental laboratory, to achieve high accuracy in detecting current air pollution for further management of environmental risks associated with road traffic.