A Machine Learning Based Method for Real-Time Queue Length Estimation Using License Plate Recognition and GPS Trajectory Data

Abstract

Real-time and accurate queue length information is crucial to developing effective queue management applications in modern traffic control systems to alleviate traffic congestion. A Random Forests (RF) based real-time queue length estimation method is proposed using the vehicle Global Position System (GPS) trajectory and License Plate Recognition (LPR) Data. The RF model is trained to predict the vehicle stop locations provided by the GPS data by features of traffic flow characteristics extracted from the LRP data. The predicted stop locations are further used to estimate the cyclic maximum queue length for each approach lane. The proposed method has been implemented on sixteen lanes of eight links from both major and minor arterials in Kunshan City, China. Key findings and conclusions include: 1) By feature selection, the travel time has the most significant impact on the prediction accuracy of the vehicle stop location, and the number of departed vehicles is the secondary informative feature. 2) The RF model achieves a satisfying accuracy for the stop location prediction and cyclic maximum queue length estimation, which has the best performance with a larger sample size in the training data. 3) Comparative analysis also shows the superiorities of the proposed model to have more accurate results by incorporating comprehensive features and a machine learning process.