High-Resolution Mapping of Vehicle Pollutant Emissions at Urban Intersections Using UAV Video Trajectory Tracking and Adaptive Microscopic Emission Modeling

Signalized intersections generate short-lived, meter-scale emission peaks through queue formation, idling, start-up, and rapid acceleration, yet conventional link- and lane-average methods suppress these localized events. This study develops an integrated unmanned aerial vehicle (UAV) video-based framework that converts aerial observations into ground-plane vehicle trajectories, refines the trajectories using an improved Iterative Adaptive Cubature Kalman Filter (IACKF), estimates second-by-second carbon monoxide (CO), hydrocarbon (HC), and nitrogen oxide (NOx) emissions using a calibrated Virginia Tech microscopic emission model (VT-Micro) with idling-state and vehicle-class corrections, and aggregates the results into 3 m × 3 m road cells. The improved IACKF combines time-decayed residual weighting, adaptive process- and measurement-noise covariance updating, and a relative convergence criterion to suppress derivative-amplified trajectory noise while preserving emission-relevant braking and acceleration transitions. Field experiments at six signalized intersections and six associated approaches in Zhengzhou achieved a mAP@0.5 (mean average precision at an intersection-over-union threshold of 0.5) of 91.7%. Controlled comparisons show that link-level averaging produces spatially homogeneous fields, lane-level averaging cannot recover longitudinal hotspot variation, and unfiltered raw trajectories generate fragmented, potentially noise-induced hotspots. In contrast, the proposed framework preserves coherent stop-line, start-up, turning, and merging patterns consistent with observed traffic operations. Quantitative visualization indicators further show that, relative to RTDC, the proposed framework reduces hotspot fragmentation density by 24.3%, increases the high-intensity color share by 11.7 percentage points, and improves spatial continuity by 1.9 percentage points. Peak-period emissions were 3.3 times the off-peak level, demonstrating the sensitivity of the mapped emission patterns to congestion-driven operating-state changes. The framework provides spatially explicit emission information for signal evaluation and sustainable traffic management.