Real-Time Prediction of Traffic Density with Deep Learning Using Computer Vision and Traffic Event Information

Abstract

Traffic congestion affects urban areas negatively in many ways. Therefore, successful and efficient traffic management is a necessity to solve or at least alleviate traffic congestion. Hence, the usage of high-quality data is not only essential but also mandatory. Prediction of traffic behaviour over a certain period of time should be done using various characteristics and related data of traffic. In this study, the location, lane, and time data of each vehicle are obtained from cameras located in D100 Highway by computer vision. Besides, the event matrices are created manually to detect the circumstances such as shoulder violation, police stop, police control, traffic flow control, weather, vehicle on the shoulder, and police or ambulance on the shoulder. The effect of the different dynamics of these events in different lanes has been transformed into a single "Event Score" with the help of the weight coefficients obtained from Logistic Regression. Afterward, the traffic density and traffic event datasets are combined to predict the next frame of the traffic. Among the many prediction algorithms tested in this study, Support Vector Machine (SVM) and Recurrent Neural Networks (RNN) were able to predict the traffic density after 1, 3, and 5 minutes with the highest accuracy. As a result of this study, it has been observed that estimation algorithms using "Event Score" obtained with separate coefficients for each lane and historical traffic density data as independent variables give successful results in dynamic and/or static traffic density estimation.