A Hybrid Traffic Forecasting Model for Urban Environments Based on Convolutional and Recurrent Neural Networks

Abstract

This article presents a hybrid model for the continuous monitoring and prediction of the spatio-temporal dependencies of road traffic. This model is built on the basis of training a YOLOv4 convolutional neural network (CNN) and two-level long-short-term memory (LSTM). The initial module of the hybrid model ensures the continuous collection of data and the filling of the database on road traffic parameters, taking into account transport infrastructure and meteorological factors. Taking the interpreted and aggregated data as input, a recurrent neural network (RNN) generates more accurate predictions. Experimental results show that the proposed model provides prediction accuracy (for a 20-minute time interval) in the range of 82–97 % using a limited number of measurements.