Vision-based vehicle emission estimation under adverse weather: integrating YOLOv11 detection with SUMO and MOVES models.

This study presents a vision-based framework for accurate vehicle speed estimation and real-world emission assessment at urban signalized intersections under challenging weather conditions. Using a publicly available dataset recorded in Curitiba, Brazil, comprising five videos (total duration 132 min) captured by a single low-cost 5 MP camera under cloudy, sunny, dusty, heavy rain, and low-visibility conditions, vehicles were detected with YOLOv11 and tracked using ByteTrack. Speed was estimated through a monocular vision pipeline calibrated against radar ground truth. The proposed speed estimation method achieved outstanding accuracy, with mean absolute error (MAE) ranging from 0.38 to 0.84 km/h and standard deviation below 0.96 km/h across all weather scenarios, significantly outperforming existing vision-only approaches. Two emission models were implemented: (1) MOVESTAR using measured speeds and vehicle classification, and (2) SUMO with real trajectory enforcement via TraCI to create a high-fidelity digital twin of the intersection. When the whole four-lane region was analyzed (439 vehicles total), MOVESTAR yielded average emissions of CO₂ 2.02 g/mi, HC 0.018 g/mi, NOx 0.058 g/mi, and fuel consumption 93.15 g/mi. After unit standardization to g/km and statistical comparison via paired t-tests, MOVESTAR consistently produced significantly lower and more realistic emission estimates than SUMO for fuel, CO₂, NOx, and HC (p < 0.01), with CO₂ and fuel showing the most significant differences (- 12 g/km and - 18.57 g/km, respectively). Heavy rain and dusty conditions were identified as the most challenging environments, whereas cloudy and sunny weather provided the highest accuracy.