SMHF: State-conditioned Mamba-hypernetwork framework for pedestrian volume estimation under observation sparsity

Providing accurate urban pedestrian volume estimates is essential for optimizing public infrastructure, improving services, and enhancing active travel experiences. However, collecting large-scale pedestrian mobility data is prohibitively expensive and often impractical in real-world scenarios. Pedestrian volume observations on road networks are typically sparse: only a small fraction of links have sensor-derived observations, while most links remain sensor-uncovered with no observations at all. Existing data-driven prediction and imputation methods rely heavily on spatially dense observations, while directly applying these models to extrapolate to numerous unobserved links may result in unreliable generalization. Moreover, prior research lacks analysis and corresponding mechanisms to adaptively adjust how transportation and land-use features (e.g., local road conditions, transit accessibility, and land-use patterns) are leveraged under observation sparsity. To address these challenges, we propose a novel State-conditioned Mamba-Hypernetwork Framework (SMHF) to estimate urban pedestrian volumes with minimal available observations. SMHF formulates estimation as a state-conditioned stepwise process: the estimator is updated at each step conditioned on an evolving estimation state sequence that dynamically fuses transportation and land-use features, graph-structured context, current pedestrian volumes with their observed/estimated status, and estimation progress. SMHF demonstrates superior performance under observation sparsity over baselines and state-of-the-art models on a real-world pedestrian volume dataset through supervised evaluation, and an independent on-site manual-count validation confirms the practical applicability of SMHF’s citywide inference on sensor-uncovered links. A sparsity-specific interpretability study provides insights into urban transportation and environmental factors associated with pedestrian activity. It also shows a consistent monotonic increase in feature-importance scores as observations become sparser, indicating a growing reliance on static features when observations are limited and highlighting the need for adaptive feature utilization under observation sparsity.