Automatic passenger counting (APC) systems underpin public transport planning and financing, yet a persistent gap separates their advertised accuracy from their field performance, with deployed systems often achieving occupancy accuracy far below vendor claims. This thesis develops a Bayesian hierarchical framework for correcting APC inaccuracy, applying it to three sensors across two Italian cities: a Wi-Fi sensor and a camera in Turin, and a Wi-Fi sensor in Asti. Each model represents the sensor error with a robust Student-t likelihood and partial pooling across stops, routes, trips, and hours, producing not a point correction alone but a calibrated predictive interval. The models are evaluated both in distribution and, critically, on held-out routes never used during training.