Please use this identifier to cite or link to this item:
|Title:||Travel time estimation on road networks : effective embedding and traffic condition representation learning||Authors:||Yang, Jingyi||Keywords:||Engineering::Computer science and engineering||Issue Date:||2021||Publisher:||Nanyang Technological University||Source:||Yang, J. (2021). Travel time estimation on road networks : effective embedding and traffic condition representation learning. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/149753||Project:||SCSE20-0474||Abstract:||Effective algorithm for travel time estimation has become increasingly important, as it is the backbone of the various services provided by urban mobility apps. With the recent advances in the field of deep learning, learning-based methods for travel time estimation have been proposed and have proved to achieve superior performance compared to traditional methods. However, two critical aspects of learning-based methods for travel time estimation remain to be further studied. First, existing methods usually use feed- forward neural network or recurrent neural network for spatial information embedding, but these two models either cannot capture the correlation among road segments or have poor efficiency. Second, representation learning of real-time traffic condition remains an area with ample space for improvement, as existing methods usually suffer from incomplete data, and overlook local traffic condition. In this report, we propose methods to address these two major concerns in deep learning-based travel time estimation. Transformer models are proposed for spatial information embedding to achieve effective encoding of road segment features and inter-segment correlations. We also propose a learned traffic map completion pipeline to address the issue of data incompleteness, and hard attention mechanism to incorporate local traffic information. Furthermore, directly conducting representation learning on unstructured traffic condition data through graph neural networks is explored. Experiments on real-world dataset show that a combination of our proposed techniques leads to a steady performance improvement compared to existing methods.||URI:||https://hdl.handle.net/10356/149753||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Updated on May 25, 2022
Updated on May 25, 2022
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.