Real-time and accurate queue length information is very imperative in evaluating the performance and to develop adequate queue management systems,especially under the congested condition in an advanced traffic control system.Many intelligent mobility technologies such as automatic number plate recognition systems and automatic vehicle identification systems for queue length estimation,have received tremendous attention due to their active deployments in recent years.A machine learning algorithm based real-time dynamic queue length estimation was proposed using the GPS and license plate recognition（LPR）data.Compared to the former shock-fitting methods,the proposed method is fully data-driven,robust,and no need for any prior knowledge or assumptions about the shockwave behaviors.In this research project,the stop locations of vehicles and 18 representative features of traffic flow characteristics around the vehicles were extracted from GPS and LPR data respectively for the training of the machine learning algorithm model.After that,a feature selection has been carried out through extracted features with the help of different feature selection techniques to remove irrelevant or redundant features,which could be harmful or at least have no contribution to the accuracy of the model.Based on the best-selected features,a Random Forest and Support Vector Regression model had been trained,and then a trained model was used to predict the stop locations of vehicles using the LPR data as input.The cyclic lane-based maximum dynamic queue lengths were estimated based on the predicted stop locations.The proposed method was implemented for thirty-nine lanes in Kunshan city,P.R China.Key findings and conclusion include:（1）By the feature selection process,the travel time in control delay feature categories had the most significant impacts on the prediction accuracy of the Random Forests model.The travel time of the leading vehicle as well as travel time of the labeled vehicle was strongly related to vehicle stop locations,for both left-turn and through lanes.Besides the travel time,the average headway of departed vehicles for both left-turn and through lanes,between two labeled vehicles in one previous cycle was also identified to have a greater contribution to model prediction accuracy.This particular feature belongs to the arrival flow category.（2）In a comparison with the performance evaluation of applied Machine Learning Models（Decision Tree,Random Forest,Linear Regression,Support Vector Regression（SVR））,the Random Forest and SVR model achieved a satisfying accuracy for the stop locations prediction with performance measures including MAE,RMSE,and MAPE.With the trained Random Forest and SVR model,the stop locations of vehicles can be successfully estimated with the collaboration of real-time LPR data.The MAE and MAPE of the lanes with the most train samples achieved the most accurate results.（3）On the basis of obtained stop locations,two lanes from thirty-nine lanes in total with satisfactory MAPE percentage scores were selected for dynamic queue length estimation phenomena.By evaluating the estimated maximum queue length on two lanes with the ground truth data,the MAE and MAPE were 13.2 meters and 14.5% for the left-turn lane and 7.2 meters and 11.9%respectively for the through lane.