Enhancing Data-Driven Predictive Modeling of Pedestrian Crowd Flow with Spatial Priors – Case Studies with Post-Event Crowd Data on a University Campus

Accurate crowd flow forecasting is crucial for effective crowd management and public safety in built environments. However, existing studies often overlook the importance of spatial connectivity in predicting crowd movements. This paper addresses this critical gap by incorporating spatial priors into crowd flow predictive modeling. We introduce a framework called the Spatio-Temporal Encoder Network (STEN), designed to simultaneously encode spatial information and temporal crowd data. Our study utilizes a newly collected dataset comprising two scenarios on a university campus, both containing spatial connectivity information. We evaluate the STEN framework’s forecasting performance, comparing it with models that rely solely on temporal crowd flow data. Results demonstrate that models whose inputs include both spatial and temporal information consistently yield more accurate prediction outcomes compared to models whose inputs are only temporal crowd flow data. This finding underscores the importance of including spatial priors in enhancing the accuracy and reliability of crowd flow predictive modeling. We envision that our study will highlight the importance of spatial connectivity for future crowd management practices, and that our STEN framework and the accompanying dataset will serve as a valuable resource for researchers developing and evaluating crowd flow prediction models and contribute to identifying potentially congested regions and providing early warnings in our built environments.