Uncertainty-Aware Bilingual License Plate Recognition with Confidence Propagation for Data-Driven Urban Traffic Forecasting

Urban transportation analytics increasingly rely on automated sensing infrastructures to generate actionable insights for planning and management. However, Automatic Vehicle Number Plate Recognition (ANPR) systems suffer performance degradation when applied to embossed license plates in developing-country environments, where illumination variability, bilingual scripts, and infrastructure idiosyncrasies introduce significant noise. This study proposes an uncertainty-aware, analytics-driven framework that integrates deep learning–based detection, bilingual optical character recognition, and structured urban intelligence modeling for end-to-end traffic analytics. We curate a bespoke dataset of 5,247 embossed Nepali license plate images captured across heterogeneous environmental conditions. Vehicle detection is performed using YOLOv8, and bilingual OCR is implemented via PaddleOCR with script-aware confidence calibration. Recognized events are aggregated to form structured event logs supporting statistical traffic density estimation and short-term forecasting using autoregressive integrated moving average (ARIMA) models with quantified uncertainty. Our empirical evaluation across five independent stratified random train-test splits shows mean detection mAP@0.5 of 96.3 ± 0.85% (daytime: 98.2%), overall recognition accuracy of 91.2 ± 1.62%, with statistically significant improvements over baseline configurations (paired t = 3.87, p = 0.0023 across all splits). Forecasting results achieve MAE = 14.3 ± 1.05 vehicles/hour and RMSE = 21.8 ± 1.42 vehicles/hour (39.4% error reduction vs. naïve persistence baseline), while prediction intervals capture 93% of observed values at 95% confidence. We further demonstrate that propagating recognition confidence into density estimation reduces aggregation bias under adverse conditions. The proposed framework extends conventional ANPR pipelines by embedding recognition uncertainty into downstream analytics, thereby providing a scalable data-driven foundation for intelligent transportation systems in resource-constrained urban environments.