Machine Learning Error Correction for Satellite-Based Virtual Tolling

Abstract

Standard GNSS positioning fails virtual tolling not because errors are large in absolute terms, but because their cost consequences are asymmetric — missed charges generate direct revenue loss while incorrect charges trigger customer disputes at substantially higher remediation cost. This mismatch is ignored by conventional correction methods, which optimize symmetric mean-squared error objectives without reference to billing economics. We present a hybrid gradient-boosting and LSTM architecture that treats lane-level positioning and boundary-crossing detection as a single coupled optimization problem, trained under an asymmetric cost objective derived from operational tolling economics. Rather than applying map geometry as a post-hoc constraint, toll-zone boundary geometry is encoded as a training feature, allowing the model to learn the statistical relationship between proximity to a boundary and the probability of a genuine crossing event. Field validation across 4,246 km of instrumented vehicle trajectories spanning urban canyon, suburban, and rural environments achieves median lateral error of 1.9 m in urban corridors against a 5.8 m baseline. Boundary-crossing detection reaches 96.3% true positive rate and 1.2% false positive rate. Satellite geometry features account for 1.2 m of the 3.9 m total urban error reduction, while HD map geometry prevents 38% of false positive boundary detections that geometric-only models cannot suppress. Real-time inference on ARM Cortex-A72 hardware averages 87 ms end-to-end latency within a 487 MB memory footprint and 2.3 W power envelope, meeting automotive-grade embedded deployment constraints without GPU acceleration.