Elsevier

Neurocomputing

Volume 432, 7 April 2021, Pages 216-226
Neurocomputing

Deep convolutional neural networks for data delivery in vehicular networks

https://doi.org/10.1016/j.neucom.2020.12.024Get rights and content

Abstract

In vehicular networks, most content delivery schemes only utilize vehicle cooperation or powerful infrastructure to satisfy data requests. How to fully utilize vehicle-to-vehicle and vehicle-to-infrastructure communications to improve data acquisition still requires further analysis. In this paper, the content delivery problem is formulated as a maximum flow of a directed network, which implies the encounters and the requests. Despite of a high delivery ratio, the proposed Content delivery scheme using mAximum Flow (CAF) is infeasible in large-scale real-time applications due to high computational complexity. To solve this problem, we transform the GPS trajectory data into two-dimensional coverage grid maps which indicate the communication opportunities between vehicles and infrastructures in CAF. The map set, which consists of coverage grid maps in a storage cycle, and the number of satisfied requests obtained from CAF compose the training set that can be trained by the deep convolutional neural networks. This solution combining CAF with deep neural networks is called CAF-Net. In the experiments, we evaluate the performances of four popular architectures of deep convolutional neural networks when outputting the targets. The results show that ResNet 50 has the smallest error and the computation time of a delivery ratio is only 82.84 ms, which is a lot shorter than 4531.53 s using CAF. The results also demonstrate the feasibility of applying the deep learning framework to vehicular networks.

Introduction

In Vehicular Ad hoc NETworks (VANETs) [1], [2], fast-moving vehicles with on-board sensors deliver contents via Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communications to support smart transportation applications, such as traffic monitoring [3] and traffic-flow prediction [4]. Currently, vehicles are equipped with Global Positioning System (GPS), radar positioning system, obstacle detection system and so on. Moreover, the fixed roadside infrastructures, such as the RoadSide Units (RSUs), have powerful processing, storage and communication capabilities [5], provide long-distance data transmissions, and connect with the content providers on the Internet [6]. Vehicles and infrastructures utilize Dedicated Short Range Communication (DSRC [7]) standard or Cellular-Vehicle-to-Everything (C-V2X [8]) technology to support V2V and V2I communications [9].

For those data acquisition services in VANETs, after a vehicular node sends content requests, other vehicles as well as RSUs may help to forward the contents. Because of the frequent change of network topology and the opportunistic communications, how to effectively utilize network resources to enhance content delivery is still an important problem. Some researches focus on vehicle cooperation, but the sparsity of vehicle distribution, high mobility and security risk of autonomous vehicles seriously restrict the performance improvement [10], [11], [12]. Our previous studies mainly use V2I communications to deliver contents by formulating the content distribution problem as bipartite graph matching problem [13], [14]. Nevertheless, integrating V2V and V2I communications thoroughly, as a complicated and open issue, provides a promising solution to data delivery optimization in VANETs. Moreover, the high computational complexity restricts the applicability of relevant algorithms, and thus how to accelerate the computation is of great significance for actual applications of VANETs.

In this paper, we propose a Content delivery scheme using mAximum Flow of a directed network (CAF), where the nodes and edges indicate the encounters between vehicles and RSUs as well as the requests from vehicles. Although CAF shows a high delivery ratio in theory and from experiments, the high computational complexity severely limits its use, especially in large-scale real-time applications. In order to accelerate the computation, we attempt to fully utilize the historical data to find the hidden relation between the V2V and V2I communication chances and the delivery ratio. Since the locations of vehicles and RSUs as well as their communication ranges can be marked on the map, the encounter information can be illustrated in images. For the regular data requests, a set of map images are relevant with the target that is the number of satisfied requests. Considering that the deep learning technology mines the pattern of images and discovers the hidden relations efficiently [15], we explore an innovative method, called CAF-Net, to calculate the metric – delivery ratio, by utilizing deep Convolutional Neural Networks (CNN). Besides deep learning, there may be other methods helpful for the computation acceleration, which require further research and detailed comparison in future work. Specifically, in CAF-Net, the trajectories of vehicles are converted to a vast quantity of coverage grid maps. Furthermore, during a certain period, a map set including several coverage grid maps and the corresponding target (the number of satisfied requests) compose the training set.

Experiments are conducted with real trajectories of taxis in Wuhan City, and eight deep CNN models are evaluated. We use the Mean Absolute Error (MAE) to measure the output difference in a model and the Root Mean Square Error (RMSE) to measure the difference between the predicted value and the observed value. The MAEs and RMSEs of all the eight models are smaller than 0.0504 and 0.0012, respectively, which show the accuracy of the proposed CAF-Net scheme. Moreover, the average times for computing a delivery ratio using the eight models are shorter than 204 ms, while CAF takes 4531.53 s. This shows the time-efficiency of these models. Among them, ResNet 50 has the smallest MAE and RMSE with the average computation time 82.84 ms. Although the training in deep learning approaches also takes some time, this work can be done off line and will not affect the overall performance.

The main contributions of this paper can be summarized as below:

  • In CAF, the data delivery problem is formulated as a classic maximum flow problem in a well-designed directed network. Theoretical and experimental results show that CAF achieves a high delivery ratio.

  • In order to utilize deep CNN models to calculate the delivery ratio, the initial GPS trajectory data is converted to a set of two-dimensional coverage grid maps, where the value of a grid implies the V2V and V2I communication chances. A set of coverage grid maps is the training image, while the number of satisfied requests in CAF is the training target.

  • In CAF-Net, eight deep CNN models based on four popular architectures are implemented, and the results show that deep CNN models have small errors and greatly shorten the computation time, which help to enlarge the application scope of CAF.

  • The idea of using deep CNN to evaluate the performance of data delivery in VANETs is innovative, and our work can be recognized as a baseline for further research.

The remainder of this paper is organized as follows. Section 2 surveys existing data delivery schemes in VANETs and the use of deep learning in transportation related domains. In Section 3, the proposed CAF scheme using maximum flow is introduced in detail. Section 4 discusses the conversion from the original trajectory data to a series of coverage grid maps, and then the construction of training set. Experiments of CAF and CAF-Net with eight deep CNN models are conducted and analyzed in Section 5 and Section 6, respectively. Eventually, Section 7 concludes this paper.

Section snippets

Related work

With the development of Internet of Things and self-driving vehicles, content downloading from the cloud and data sharing between end-devices attract more and more attentions [16], [17]. Hadaller et al. use roadside access points to provide Internet access for vehicles via IEEE 802.11, and report on the data gathered by four capture devices from vehicles driving on a rural highway. They demonstrate that vehicular opportunistic access can be greatly improved if the wireless conditions in the

CAF scheme

In VANETs, fast-moving vehicles send data requests to nearby RSUs. Then, RSUs download some data from the content providers on the cloud. Thereafter, the vehicles obtain contents through V2I and V2V communications. In order to satisfy as many requests as possible, in the proposed CAF scheme, the data delivery problem is formulated as a classical maximum flow problem. A content delivery solution using Edmonds-Karp algorithm [38] to achieve the high delivery ratio is obtained. Here the delivery

Traffic dataset

In general, the deep CNN approach requires a large number of pixels and samples in its training set,while the trajectory data in Sanya [39] that is used to test the performance of CAF in Section 5 is not large enough for further analysis. Hence, we utilize the trajectory dataset of taxis in Wuhan, Hubei province, China for six days [40]. Considering that it is down sampled data, which is collected every 5 to 10 s, the trajectory of a vehicle is usually incomplete and only a small number of

Experiments of CAF

For the data delivery research in vehicular networks, there are a small number of public standard datasets and different datasets have different features. In this paper, we use two public trajectory datasets of taxis, which are collected in Sanya, Hainan province, China [39] and Wuhan, Hubei province, China [40], respectively. The trajectory dataset in Sanya is collected every second during one hour, while the dataset in Wuhan is collected every five to ten seconds during six days. Thus, the

Dataset

Based on the training set given above, a large number of map sets like Fig. 6 are generated from the taxi trajectory in Wuhan, as well as the training targets numsat. The statistics of the training set is shown in Fig. 9. From Fig. 9a, a large majority of delivery ratios are between 0.55 and 0.75. As illustrated in Fig. 9b, during almost all the storage cycles, the number of requests in the whole network falls within [550,650]. Therefore, most values of the training target reqsat are between

Conclusion

For data acquisition services in VANETs, we propose an approach using deep convolutional neural networks for data delivery, named CAF-Net. In the data transmission scheme CAF, the data delivery problem is formulated as a maximum flow of a directed network, where the flow implies the delivery ratio. Furthermore, deep CNN models are utilized to enhance the computation efficiency. The taxi trajectory during a period of time is mapped into a coverage grid map, and a large quantity of map sets and

CRediT authorship contribution statement

Hejun Jiang: Conceptualization, Methodology, Software. Xiaolan Tang: Conceptualization, Writing - original draft, Supervision. Kai Jin: Software, Investigation. Wenlong Chen: Validation, Supervision. Juhua Pu: Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

We gratefully acknowledge the support from the Beijing Natural Science Foundation (4202012), the National Natural Science Foundation of China (61872252), the National Key R&D Program of China (2017YFC0803700), and Science & Technology Project of Beijing Municipal Commission of Education in China (KM201810028017).

Hejun Jiang received the B.E. degree in computer science and technology with the College of Information Engineering, Capital Normal University, Beijing, China, in 2016. He received the M.S. degree with the College of Information Engineering, Capital Normal University, in 2019. His research interests include vehicular networks, urban computing and machine learning.

References (46)

  • J.A. Dias et al.

    Performance evaluation of a cooperative reputation system for vehicular delay-tolerant networks

    EURASIP Journal on Wireless Communications and Networking

    (2014)
  • M.A. Salahuddin et al.

    Software-defined networking for RSU clouds in support of the internet of vehicles

    IEEE Internet of Things Journal

    (2015)
  • W. Xiong et al.

    Measurement and characterization of link quality for IEEE 802.15.4-compliant wireless sensor networks in vehicular communications

    IEEE Transactions on Industrial Informatics

    (2016)
  • X. Wu et al.

    Vehicular communications using DSRC: challenges, enhancements, and evolution

    IEEE Journal on Selected Areas in Communications

    (2013)
  • M. Gonzalez-Martin et al.

    Analytical models of the performance of C-V2X mode 4 vehicular communications

    IEEE Transactions on Vehicular Technology

    (2019)
  • F. Jameel et al.

    Performance analysis of cooperative V2V and V2I communications under correlated fading

    IEEE Transactions on Intelligent Transportation Systems

    (2019)
  • W. Li et al.

    ART: An attack-resistant trust management scheme for securing vehicular ad hoc networks

    IEEE Transactions on Intelligent Transportation Systems

    (2016)
  • X. Tang et al.

    Distributed storage scheme using bipartite graph matching for vehicular networks

    Journal of Software

    (2016)
  • X. Tang et al.

    Data acquisition based on stable matching of bipartite graph in cooperative vehicle-infrastructure systems

    Sensors

    (2017)
  • P. Bellavista et al.

    Towards better scalability for IoT-cloud interactions via combined exploitation of MQTT and CoAP

  • D. Hadaller, S. Keshav, T. Brecht, S. Agarwal, Vehicular opportunistic communication under the microscope, in: ACM...
  • X. Tang et al.

    Data dissemination based on fuzzy logic and network coding in vehicular networks

    Wireless Communications and Mobile Computing

    (2017)
  • F. Malandrino et al.

    Optimal content downloading in vehicular networks

    IEEE Transactions on Mobile Computing

    (2013)
  • Cited by (0)

    Hejun Jiang received the B.E. degree in computer science and technology with the College of Information Engineering, Capital Normal University, Beijing, China, in 2016. He received the M.S. degree with the College of Information Engineering, Capital Normal University, in 2019. His research interests include vehicular networks, urban computing and machine learning.

    Xiaolan Tang received the Ph.D. degree in computer application technology with the School of Computer Science and Engineering, Beihang University, Beijing, China, in 2014. She is currently an Associate Professor with the College of Information Engineering, Capital Normal University, Beijing, China. During August 2016 and August 2017, she was a Visiting Scholar with Harvard Medical School, Boston, MA, USA. Her research interests include vehicular networks, wireless sensor networks, urban computing, and driving behavior analysis.

    Kai Jin received the B.E. degree in computer science and technology, College of Information Engineering, Capital Normal University, Beijing, China, in 2016. He is currently a research associate in the College of Information Engineering, Capital Normal University, Beijing, China. His research interests include face recognition, facial expression analysis, and deep learning. He was a recipient of the First-Prize National Scholarship from the Ministry of Education of China in 2016.

    Wenlong Chen received the Ph.D. degree in communication and information system from University of Science and Technology Beijing, Beijing, China, in 2011. He is currently an Associate Professor with the College of Information Engineering, Capital Normal University, Beijing, China. His research interests include network protocol, Internet architecture, high performance router and wireless sensor networks.

    Juhua Pu received the Ph.D. degree in computer application technology with the School of Computer Science and Engineering, Beihang University, Beijing, China, in 2005. She is currently an Associate Professor with the School of Computer Science and Engineering, Beihang University. During April–July 2004, she was a Visiting Scholar with the Hong Kong University of Science and Technology, Hong Kong, China. During January 2011 and January 2012, she was a Visiting Scholar with Georgia Institute of Technology, Atlanta, GA, USA. Her research interests include urban computing, vehicular networks, and smart cities.

    View full text