Analysis of Highway Traffic Using Deep Learning Techniques / Derin Öğrenme Teknikleri Kullanılarak Anayol Trafik Analizi

Abstract

Traffic flow forecasting has an important place in designing a successful intelligent transportation system. The success of the forecasting is related to the accuracy and timely acquisition of the traffic flow data. The inadequacy in the number of data has led to the use of shallow architectures in the traffic forecasting models realized so far or to design models with generated data. These models failed to occur forecast results with sufficient success. Nowadays, in the age of big data, in parallel with the increase in traffic density, there has been a significant increase in the diversity and size of the collected traffic flow data. This result constitutes the main motivation in our study. Our study aims to forecast traffic density at the exit of a motorway that have linked roads. The forecasting models proposed in our study were designed using generally accepted, Deep Learning techniques, which can occur meaningful prediction results with big data. The techniques used in our study are Recurrent Neural Network (RNN), Long-Short Time Memory (LSTM), Stacked Long Short-Term Memory (S-LSTM), Bidirectional Long Short-Term Memory (B-LSTM) and Gated Recurrent Unit (GRU) neural networks. The dataset used in the study consists of 929 thousand 640 measurement data collected by loop sensors placed at 6 different points. Three different training data sets were created, split 90%, 80% and 70% of all data and the remainder of the data used as the test dataset. Forecast achievements of the designed models on the test dataset were recorded by calculating the Mean Square Error (MSE) values. In addition, all models are run with different number of epochs and the effect of the training set size and iterations on learning was investigated. The results show that Deep Learning techniques in traffic flow forecasting with low MSE values occur successful results and can be used in traffic flow prediction models. When the results of selected Deep Learning techniques and designed models are compared, it is observed that B-LSTM has the best forecast performance with the lowest MSE value of 36,60.