| With the acceleration of urbanization and the complexity of urban functional facilities,how to effectively and intelligently optimize the daily travel experience of residents is of great practical significance for the construction of people-oriented smart cities and intelligent transportation systems.Among them,accurate travel time estimation has become the primary task of optimizing residents’ travel,and it is the core component of downstream tasks including route navigation,traffic control,and unmanned driving.Thanks to the development of urban big data,the research on travel time estimation is also tending to be refined and high-precision.However,the crowd-sourced,spontaneous and uncontrollable nature of big data will inevitably lead to problems such as heterogeneity and sparseness..Most of the existing travel time estimation studies have not considered the effective learning and fusion of heterogeneous sparse data.Therefore,starting from the learning and fusion of spatiotemporal features of heterogeneous sparse data of urban traffic environment,this paper deeply understands urban traffic environment and traffic elements at the feature correlation level,so as to improve the recognition and learning of urban potential patterns in travel time estimation task.The main research contents of this paper are:(1)Cognition of travel time estimation research in urban traffic environment: First,the representative methods of existing research are summarized from the development and classification of travel time estimation research at home and abroad,and the abstract modeling of urban traffic environment from the perspective of existing research The essential cognition of the existing research on the urban traffic environment is analyzed.And further based on the literature,the existing research on the classification and utilization of traffic elements is summarized.(2)Spatio-temporal feature learning and fusion of urban heterogeneous sparse data: Starting from the path elements in the urban traffic environment,a large-scale road network representation learning framework is proposed that takes into account the movement patterns of urban vehicles,coupling the inherent topology and load carrying of urban road networks.In order to learn a more effective representation vector of urban road network based on the vehicle movement pattern based on the model,starting from the global elements in the traffic environment,a spatiotemporal feature learning and fusion method considering the correlation of global elements is proposed,and the correlation relationship and representation method of global elements are analyzed from the feature level.(3)Actual case verification: on the actual data sets of Shenzhen and Xi’an,the progressively increasing travel time estimation model is used,and the path elements and the global elements are selected to represent the learning method scheme and set up comparative experiments to verify the effectiveness of the method in this paper.Starting from ETA research,this paper solves the problem of learning and fusion of heterogeneous and sparse urban traffic elements in current research,proposes a more effective representation method for urban traffic elements,and verifies the effectiveness and applicability of this method in practical cases. |
