Sensing and Predicting Urban Rail Platform Crowding Using Emerging Data Sources

Abstract

Rail platform crowding poses serious challenges to passenger safety, operational performance, and service quality in urban rail transit systems. This thesis investigates the short-term forecasting of platform-level crowding, focusing on enhancing prediction accuracy, spatial granularity, and operational interpretability through multi-source data integration. We first employ a gradient-boosted tree regression model (LightGBM) to leverage fare card transaction, vehicle location, weather, and public event data from the Washington Metropolitan Area Transit Authority (WMATA) to forecast platform-level occupancies 15–60 minutes ahead of time. Our results show significant improvements over a WMATA-internal baseline while providing a robust data preparation and prediction pipeline. Subsequently, we explore integrating platform-level CCTV data to overcome the lack of real-time crowding estimates. Using a custom-collected image dataset and three computer vision methods, namely object detection (YOLOv11, RT-DETRv2) and head counting (APGCC), crowd-level classification (Crowd-ViT), and semantic image segmentation (DeepLabV3), we demonstrate that estimated counts from calibrated image segmentation maps enable accurate real-time estimation of platform crowding. Additionally, we show that these estimates can correct and improve 15-minute horizon predictions when incorporated with a stochastic gradient-boosted tree learner such as LightGBMLSS. Finally, we extend the time series modeling framework by incorporating network-wide causal influences through an analysis driven by Empirical Dynamic Modeling and Convergent Cross Mapping. We show that accounting for network effects improves predictive performance, particularly for platforms characterized by regular low-occupancy patterns, improving the prediction of anomalies. The work presented in this thesis extends the existing literature on short-term platform crowding prediction, offering new methodologies to incorporate emerging CCTV data and causal network effects for increased prediction accuracy.