Multi-Class Traffic Density Forecasting in IoV using Spatio-Temporal Graph Neural Networks

Abstract

Internet of Vehicles (IoV) is an emerging archetype that is a distributed network of various vehicles armed with sensors, actuators, technologies, and applications to connect and exchange data with each other over the Internet. The primary goal of IoV is to provide a vehicular platform to enable better communication and Quality of Service (QoS) for vehicles, pedestrians, and roadside infrastructure in real-time, through the use of Vehicle-to-Vehicle (V2V), Vehicle-to-Pedestrian (V2P), Vehicle-to-Infrastructure (V2I), Vehicle-to-Network (V2N), and Vehicle-to-Cloud (V2C) channels. However, the increasing number of vehicular services poses serious concerns and challenges to the researchers: real-time traffic forecasting, service placement, security, reliability, and routing. The primary focus of this work is concerned with the challenge of multi-regional forecasting of multi-class traffic. These traffic forecasting models enable pro-activeness in systems by providing real-time and accurate predictions. Also, they can explore traffic densities over spatial and temporal domains for various vehicle types. However, current literature cannot provide forecasts for multi-region and multi-class vehicles at the same time. This study aims at enabling a proactive platform which could make decisions based on the integrated Graph Neural Network (GNN) and Gated Recurrent Unit (GRU) based traffic forecasting model, i.e., Spatio-Temporal GNN (STGNN) based traffic forecasting model. The STGNN-based traffic forecasting model uses the GNN and GRU models to explore spatial and temporal features of varying vehicular multi-class traffic densities. In GNN, spatial data consisting of multi-class traffic densities are utilized for the feature extraction that results in graph embeddings. In the GRU, these graph embeddings are utilized for temporal feature extraction. This approach enables the forecasting of multi-class vehicular traffic densities and the pro-activeness of an IoV platform. In addition, the performance results show that an intelligent platform can be built upon the proposed traffic forecasting model that is capable of inspecting complex and nonlinear traffic accurately.