| Resumo: | Nowadays, the technology to turn cities smart already exists. Smart Cities, as they are called, are capable to sense, analyze and react: sense through the set of sensors displaced along the city, as they are sensors either xed (for environmental monitoring) or moving (for instance, citizens with their smartphones). A notable case is Porto, which incorporates a mesh network with more than 600 vehicles (buses, taxis and garbage trucks), communicating in-between and enabling the passengers of the buses of the city major bus carrier to access freely to the Internet while commuting. A vehicular network like this has huge positive impact in the city mobility, which is one of the biggest concerns of the governmental institutions. Therefore, it is crucial to understand what can be done to improve mobility. By analyzing the data generated by the movement of the buses, it is possible to deliver a new set of tools that might be useful for the everyday life of the bus passengers and bus eet managers. From the passengers perspective, the utility can be brought by the introduction of smart schedules, which consists on delivering estimated time of arrival that is adapting itself to the city dynamics, through the evolution of the time, and that can be accessed directly from their smartphones. From the perspective of the bus eet managers, it is possible to deliver insights about the usual behaviour of their bus lines, giving openness for them to react to the new or abnormal city public transportation dynamics. This dissertation presents an approach for analyzing the data descendent from the vehicular network and how to use it to answer the previously addressed problems. Regarding the missing link between the GPS trace from the bus and the bus line that they are doing, a map-matching algorithm is implemented. That turns possible the computation of estimations and predictions of the bus' passing times. In what concerns prediction, three machine learning ensemble algorithms have been tested. Finally, proof-ofconcept applications are implemented to demonstrate the real-life applicability, by helping the bus passengers and bus eet managers to react to the di erent events of their quotidian. The results show that the map-matching algorithm presents a good quality. Also, they demonstrate that the best machine learning algorithm, considering the prediction error, is Bagging using Support Vector Regressor as the base estimator. Finally, the pro les obtained in the performance dashboard enable distinction between optimal and non-optimal bus lines.
|
|---|