A machine learning-based overlay technique for improving the mechanism of road traffic prediction using global positioning system

Abstract

Global Positioning System (GPS)-based road traffic prediction is one of the predominating technologies in the modern technological era, which facilitates smooth navigation and reduces mobility time. Various Navigation Maps are used worldwide for traffic congestion and delay prediction, which relies upon the GPS location of the individual’s smartphone to predict traffic congestion and delay utilizing stored data and current GPS locations. However, this method sometimes malfunctions due to the uneven distribution of passengers in different vehicle types on the roadway as there are far more passengers in buses as compared with trucks, if few buses are present in the traffic stream then it will show congestion and delay in traffic. Existing mapping techniques possess limits to incorporate the classified vehicle count and categories of vehicles. To mitigate such limitations, this work overlays the information of GPS localization, using existing Maps, with classified vehicle count and vehicle categories to estimate better road traffic congestion and delay. We consider two mid-sized Indian cities in the state of Uttar Pradesh (Varanasi and Gorakhpur) due to the diverse nature of mixed road traffic. For classified vehicle count data, video recording was carried out by using video recording cameras at various sites in both considered cities. Next, different handcrafted features are extracted from the collected traffic volume data prior to the training of the machine learning-based forecasting models (ARIMA and SVM) to predict traffic volume. The classified road traffic vehicle utilizes previously observed values for prediction, thereby helps in making a good decision about route selection and traffic management. Further, this work annotates the forecasted data overlay with GPS value as per the traffic condition to build an XGBoost-based classification model. Finally, the results demonstrated the effectiveness of the proposed technique and highlighted the importance of integrating classified vehicle count and categories of vehicles with GPS. We achieve forecasting accuracy of more than 93% for both ARIMA and SVM forecasting models, followed by more than 95% accuracy of prediction via XGBoost. © Springer Nature Switzerland AG 2024.