| Summary: | In recent times, governments around the world have increasingly focused on transforming their cities into smart cities. A large number of these smart city projects requires the collection and analysis of datasets on a large scale. These are conducted by the various entities of the society, e.g., government organizations or
businesses. To date, numerous data-driven computational techniques have been proposed to perform such analysis, and to draw previously untapped insights from the datasets. This dissertation focuses on such techniques in the domain of urban mobility of smart cities. Specifically, it introduces three problems related to the estimation of travel time and demand of vehicles, which play critical roles in urban mobility. From the perspectives of transport operators, e.g., taxi or ride-hailing companies, such information encourages a more effective allocation of existing resources or planning of routes, which in turn improves commuter experience. For public organizations, such information helps in the understanding of the traffic conditions within the transportation network, which in turn supports urban planning. Finally, such information also helps commuters to understand the travel time of routes they take and plan their trips more effectively.
In the first problem, we recognize that existing work for the estimation of the travel time of trip components within the metro network either requires datasets that are usually hard to obtain, or considers the problem on scales too small (e.g., consider only single-line scenario). In some cases, there are also limitations in terms of computational efficiency. Due to the inaccessibility to fine-grained datasets like cell-phone datsets, the smart card data is a natural choice for such estimations. However, what makes such estimations challenging is that smart card data only provides information about the origins and destinations, but not the actual routes
taken between them. Since there may often be multiple route choices between an origin and destination, it is not straightforward to know what exact routes were taken, and thus challenging to estimate the trip component times. To tackle these limitations, we propose a novel multi-stage approach that accounts for the issue of multiple route choices within the network, while leveraging solely the smartcard dataset to infer these component times. The key idea is that we first seek to identify the exact routes taken by passengers and the travel time associated with it for selected trips, using insights to the metro network and approaches like trip time distribution fittings. Then, we try to estimate the time variables of trip components by considering linear formulations of the components of these trips, as well as other issues like overfitting and robustness of estimation.
In the second problem, we note that existing methods for path travel time estimation generally require single-source large datasets for a satisfactory accuracy, and may not be suitable for vehicles with small datasets. Furthermore, considering only single-source datasets neglects the estimation-improving insights available
in other datasets. A challenge is how to effectively utilize such other trajectory sources. This challenge is also further compounded by that the path travel time is generally dependent on various temporal and spatial factors. As such, we propose Attribute-related Hybrid Trajectories Network (AtHy-TNet), the first neural model
that effectively considers the relative importance of different attributes under different scenarios, as well as the spatial and temporal relationships across hybrid trajectory data, to estimate the path travel times. Not only does this model consider the various dependencies affecting travel time, it also models the multi-nature
correlations across trajectories that may come of different vehicle types.
For the third problem, we recognize that existing work for regional vehicle demand prediction largely depends on the correlations of historical demands, which may not be accurate in some situations. As such, guiding the estimation via highlevel attributes like period of day may help to resolve such inaccuracies. However, the state of the art largely ignores these high-level attributes and their relative importance under different circumstances. Furthermore, while the state of the art considers pair-wise relationships across different regions, it does not consider the mutual correlations across these relationships. Similar to path travel time
estimation, what makes this problem challenging is that there exists many multi-nature dependencies for demand of vehicles. As such, we propose Attribute-guided Multi-graph Convolutional Network (AM-Net), a novel neural model that models the spatiotemporal dependencies of regional demands via three different pair-wise relationships and their mutual correlations, and guides the estimation using high-level temporal and spatial attributes.For each of the three problems, we evaluate the proposed solution with real-world datasets, and demonstrate its superior performance over the corresponding state of the art. We also conduct in-depth analysis on how the method performs under different conditions, e.g., period of day and day of week.
|
|---|