Research On Learning Based Lane Keeping And Lane Changing Behaviors Of Intelligent Vehicle

Lane keeping and lane changing are two common driving behaviors in vehicle driving.The fundamental purpose of studying lane-keeping behavior is to control the driving direction of the vehicle and make the vehicle drive on the current Lane automatically,so as to reduce the driver's operating burden and improve driving safety and comfort.The purpose of studying lane changing behavior is to enable vehicle to make decision and execution of lane changing reasonably and orderly in specific driving scenarios(especially when traffic flow is large and frequent lane-changing occurs),so as to relieve traffic flow,improve road capacity and alleviate traffic congestion to a certain extent.Intelligent vehicles differ from traditional vehicles in that they have the functions of environment perception,planning and decision-making,and motion control,and can perform the expected driving behavior in an intelligent way.The modeling analysis and research of lane-keeping and lane changing behaviors of intelligent vehicles have wide application value in the fields of adaptive cruise control,intelligent auxiliary driving and autonomous driving.The traditional lane keeping model and lane changing model are mostly designed based on rules,which have two problems.On the one hand,the driving scenarios are often complex,and it is difficult for the traditional model to take all kinds of situations into account.On the other hand,the traditional lane keeping model and lane changing model do not take into account the driver's behavior habits,which makes it difficult for the model to effectively reflect the inconsistency and nondeterminacy of a series of psychological and physiological activities such as driver perception,decision-making and execution.Because the learning-based lane keeping and lane changing model are data-driven,the model can self-explore the characteristics hidden in the data,and then use the extracted features to drive the vehicle to perform the desired driving behavior.From the perspective of data-driven,this paper adopt learning-based method to study the lane keeping and lane changing behaviors of vehicles on the highway.The specific research contents are as follows:First,learning-based lane-keeping is studied.A lane keeping model based on end-to-end deep learning is designed for two-lane simple driving scenarios.The validity of the model is evaluated on a large number of data samples collected manually,and the interpretability of the model is improved by visualization.The validity of the model is evaluated on large-volume data samples collected manually,and the visualization method is used to improve the interpretability of the model.In addition,based on the end-to-end deep learning,a data aggregation algorithm is used to design a lane keeping model based on imitation learning.The model assists the vehicle in completing the lane keeping task using an expert sample generated by interaction with the environment.Second,learning-based lane changing decision is studied.For the complex driving environment including multi-vehicle and multi-lane,different learning methods(Support Vector Machine,Multi-Layer Perceptron network and Random Forest and Gradient Boosting Decision Tree in ensemble learning)are used to establish the autonomous lane changing decision model.The predicted results of the model are compared comprehensively from six standard of analysis: accuracy,precision,recall,F1-measure,AUC and Kappa coefficient.Third,learning-based lane changing execution is studied.Considering the time dependence of the whole lane changing process in multi-vehicle and dual-lane driving scenarios,a lane-changing execution model based on Long Short-Term Memory(LSTM)network is designed to predict the trajectory of vehicle lane changing execution,and the prediction results are compared with those of Multi-Layer Perceptron(MLP)network.