LE RETI NEURALI ARTIFICIALI PER LA STIMA DEI FLUSSI DI TRAFFICO NELL’AMBITO DEI SISTEMI DI MONITORAGGIO

This doctoral thesis, titled "Artificial Neural Networks for Traffic Flow Estimation in Monitoring Systems," investigates the application of Artificial Neural Networks (ANNs) to enhance mobility flow monitoring within Intelligent Transportation Systems (ITS). The primary objective is to evaluate the ability of ANNs to perform spatial extension of data—specifically, to estimate traffic or passenger flows on unmonitored links of a network based on real-time data collected from a limited number of physical sensors or monitoring stations. While ANNs are widely used for short-term temporal forecasting in transportation, their use for spatial extension remains relatively unexplored in literature. This research addresses this gap by proposing a methodology where ANNs are trained using supervised learning on datasets generated through advanced traffic simulation models. This approach allows the system to overcome the lack of real-world data for unmonitored links. The thesis presents two main real-world case studies:  Urban Road Traffic: Tested on the road network of Benevento, Italy. The results demonstrate that ANNs can accurately estimate flows on unmonitored streets, achieving high coefficients of determination (R2 ) and low error rates, especially when existing demand patterns (OriginDestination matrices) are incorporated into the training process.  Subway Passenger Flows: Applied to Line 1 of the Naples Metro, Italy. The model successfully predicts passenger loads between stations using only entrance turnstile data, providing a cost-effective alternative to expensive on-board infrared or weight sensors. The findings confirm that ANNs are a powerful and computationally efficient tool for traffic monitoring. They enable operators to obtain comprehensive network-wide information without the high costs associated with installing and maintaining physical sensors on every single link. Future research directions include testing multi-layer deep learning architectures and integrating spatial extension with short-term temporal forecasting