Spatial-Temporal Indoor Crowd Interpolation and Prediction with Graph Neural Networks

Abstract: Crowd analysis is crucial for smart city applications, such as space management, safety, and well-being. While studies address spatial, temporal, and spatial-temporal dimensions, spatial-temporal data granularity is often constrained by the high cost and maintenance of high-resolution sensor deployment. Existing approaches focus heavily on prediction and rely on predefined graph structures, overlooking dynamic spatial-temporal dependencies and underperforming in interpolation tasks. To address this, we propose a hybrid model, the Spatial-Temporal Attention-based Interpolation and Prediction Graph Convolutional Gated Recurrent Unit (STAIP-GCGRU), which integrates interpolation and prediction tasks to deliver robust spatial-temporal predictions, even with incomplete graph structures.