ED-ACNN: Novel attention convolutional neural network based on encoder–decoder framework for human traffic prediction

Accurate human traffic prediction, as a vital component of an intelligent transportation system (ITS), can not only reduce traffic congestion and resource consumption, but also provide a foundation for other tasks, such as risk assessment and public safety. Owing to the rapid development of computin...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Applied soft computing Jg. 97; S. 106688
Hauptverfasser: Pu, Bin, Liu, Yuan, Zhu, Ningbo, Li, Kenli, Li, Keqin
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier B.V 01.12.2020
Schlagworte:
Attention mechanism
Encoder–decoder framework
Spatio-temporal data mining
Traffic-flow prediction
Attention mechanism
Encoder–decoder framework
Spatio-temporal data mining
Traffic-flow prediction
ISSN:1568-4946, 1872-9681
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Accurate human traffic prediction, as a vital component of an intelligent transportation system (ITS), can not only reduce traffic congestion and resource consumption, but also provide a foundation for other tasks, such as risk assessment and public safety. Owing to the rapid development of computing power, massive data storage, and parallelization, deep-learning techniques, especially convolutional neural networks (CNNs), have become a powerful tool for traffic-flow forecasting. However, most of these methods in the literature over-emphasize the accuracy of traffic-flow forecasting and ignore its efficiency. It is often beneficial to develop smaller models (e.g., fewer model parameters) to improve efficiency. In this work, taking into account the efficiency and accuracy of the prediction, a novel attention CNN based on an encoder–decoder framework, called ED-ACNN, is proposed. First, the convolutional layer is considered the coding layer to extract spatial and temporal correlations. Then, the deconvolution layer as a decoding layer is expertly designed to reconstruct the future traffic-flow image. Next, the attention mechanism is introduced into the proposed model to capture the correlation between the spatial traffic-flow images’ channels. Finally, for the three characteristics of closeness, period, and trend, it is concluded that the closeness feature is the most significant for human traffic prediction in the proposed approach. An extensive experimental evaluation of two types of real-world crowd flow (Beijing and New York City) is presented, and the results show that the proposed method can be very competitive with state-of-the-art baselines. [Display omitted] •ED-ACNN is proposed to improve accuracy and efficiency for traffic flow data.•Two types of fusion of traffic flow data are presented for spatio-temporal forecasting.•Three attributes of traffic flow are analyzed, and which one is most important is listed.•Experiments verify the performance on two real-world traffic flow datasets.
ISSN:1568-4946
1872-9681
DOI:10.1016/j.asoc.2020.106688