Deep convolutional neural networks for data delivery in vehicular networks

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 out putting 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. (c) 2020 Elsevier B.V. All rights reserved.

Automated summary

In this paper, a solution combining maximum flow and deep neural networks, called CAF-Net, is proposed for the content delivery problem in vehicular networks. Experimental results show that ResNet 50 has the smallest error and significantly reduces the computation time for delivery ratio, demonstrating the feasibility of applying deep learning to vehicular networks.