A framework for vehicle tracking and fine-grained spatio-temporal statistical load estimation using an infrared traffic camera

Heavy-vehicle traffic accelerates pavement deterioration and increases lifecycle costs, motivating the need for scalable load estimation. Weigh-in-motion (WIM) stations measure axle loads but are expensive to deploy at scale. This work proposes an infrared-camera-based, scalable WIM proxy for traffic loading by transferring WIM-informed load priors to locations without embedded sensors. Vehicles are localized in 3D using instance segmentation and projective geometry, while axle configurations are inferred from thermal wheel signatures. Axle loads are estimated by mapping inferred configurations to WIM-derived load distributions, and multi-frame tracking propagates these estimates along vehicle trajectories to characterize spatio-temporal loading patterns. Evaluation against synchronized WIM records yields a 33.7% MAPE for axle-load estimation and 84% axle-count prediction accuracy. Contact patch centre localization achieves a 0.44 m MAE relative to reference contact patches derived from WIM. Comparison with observed pavement condition shows that the estimated loading patterns correlate with deterioration, supporting cost-effective, privacy-preserving, network-scale infrastructure monitoring.