Utilizing Random Forest and Neural Network to Extract Lane Change Events on Shanghai Highway

This paper considers a comprehensive naturalistic driving experiment to collect driving data about vehicle’s lane change behaviors on an actual Shanghai highway. Using the acquired real-world natural driving data, a lane-change database with kinds of vehicles’ driving information is established. A novel way of selecting variables is proposed by considering the values and changes of various variables over time periods. The random forest (RF) and Gini coefficient are adopted for variables’ importance ranking, and a total of seven top contributing factors are selected. The results indicate that the variables concerning vehicles’ acceleration and speed, especially the standard deviation of longitudinal acceleration, exerts the greatest influence. Subsequently, based on these key factors, the long short term memory (LSTM) is utilized for performing training and predicting operations on our collected naturalistic driving data. Finally, good effects and high accuracy on the recognition and extraction of lane change behavior are obtained.