Vehicle Intersections Prediction Based on Markov Model with Variable Weight Optimization

Abstract

In this study, a new algorithm for predicting vehicle turning at intersections is proposed. The method is based on the Markov chain and can predict vehicle trajectories using GPS location sequences. Unlike traditional Markov models, which use preset weights, we created the Markov model using a data-driven weight selection method. The proposed model can dynamically adjust the weights of each intersection’s influence on current trajectories based on the data, in contrast to the fixed weights in traditional models. The study also details how to process trajectory data to identify whether a vehicle has passed through an intersection and how to determine the adjacency relationship of intersections, thus providing a reference for implementing a model of the classification problem. The data-driven algorithm was applied and compared to the fixed-weight algorithm on the same trajectory dataset, and the superiority of the weight selection algorithm was proven. The prediction accuracy of the traditional method was 49.61%, while the proposed method achieved a prediction accuracy of 60.66% for 100,000 trajectory datasets, nearly an 11% increase. Volunteer participation in the second dataset collected on the university campus showed that the accuracy of the proposed method could be further improved to 79.31% as the GPS sampling frequency increased. Simulation results show that the algorithm provides accurate prediction and that the prediction effect is improved with the expansion of the trajectory data set and the increase in GPS sampling frequency. The proposed algorithm has the potential to provide a location-based optimization of network resource allocation.