Advancing crowd forecasting with graphs across microscopic trajectory to macroscopic dynamics

The high-density multi-directional passenger crowd within large transportation hubs raises practical concerns related to degraded flow conditions and possible safety hazards, but also represents a challenge to mainstream crowd dynamic forecasting methods for several reasons: they involve the dense and heterogeneous-destination passenger crowd, which are hardly studied compared to their diluted or homogeneous counterparts, in a complex context, made of four-directional pedestrian intersections. In light of the need for real-time, safety, and efficiency-oriented management in crowded scenarios, we introduce a Graph Neural Network-based Crowd Forecaster (GCF) designed to forecast crowd evolution across three dimensions: (i) the individual's trajectory at the microscopic level; (ii) ‘sub-regional’ safety and efficiency, gauged by Crowd Danger (Cd) and Passing Distance (Pd), at the mesoscopic level; and (iii) the collective dynamics of the crowd within the ‘global region’, depicted through the fundamental relationship, at the macroscopic level. Comparing both classic and state-of-the-art models across physics-based, learning-based (i.e., sequence-learning and structure-learning) categories for their forecasting performance, the outcomes reveal: (i) our GCF model exceeds others in both individual, ‘sub-regional’ and ‘global’ scales, indicating its potential for real-time crowd intervention; (ii) GCF demonstrably upholds the recognized strengths of physics-based (aptitude for dense crowd) and learning-based methods (trajectory prediction precision); (iii) the necessity to encompass predictions of mesoscopic and macroscopic features, rather than solely focusing on trajectories, is underscored by BiLSTM's subpar performance in these aspects, despite its relative advantage in forecasting individual's trajectory, thereby endorsing the multi-dimensional forecasting approach this paper advocates.