Research On Active Lane-Changing Decision And Trajectory Planning Of Autonomous Vehicle Based On Deep Learning

Automatic lane-changing is an important scenario in smart car driving.Unreasonable lane-changing is often an important cause of traffic congestion and traffic accidents.Aiming at the problems of insufficient research on nonlinear characteristics of autonomous vehicles active lane-changing decision-making under dynamic traffic environment,low decision-making accuracy,strong dependence of lane-changing trajectory planning algorithm on road shape and low safety,this paper conducts research on autonomous vehicle active lane-changing decision-making and lane-changing trajectory planning methods under dynamics environment based on deep learning.The study has conducted in-depth research on the nonlinear characteristics in lane-changing decision-making,and improves the versatility and safety of the lane-changing trajectory method.The main research contents of this paper are as follows:(1)In order to ascertain the importance of each influencing factor to the lanechanging decision,the XGBoost algorithm is used to quantify the importance of each influencing factor in the lane-changing decision,and the importance of each characteristic variable in the lane-changing decision is explored.Use next-generation simulation(NGSIM)data to construct a lane-changing decision dataset,because the NGSIM data contains a lot of noise information,the local weighted scatter smoothing(LOWESS)algorithm is used to filter the position,velocity,and acceleration,and then the lanechanging decision variables are extracted,and finally the lane-changing decision dataset is constructed.(2)Use deep learning’s powerful extraction and fitting capabilities for nonlinear features,Propose a lane-changing decision model based on a variant of recurrent neural network(LSTM)and gated recurrent unit(GRU),and combine it with lane-changing model based on backpropagation(BP)neural network and traditional machine learning algorithm-support vector machine(SVM)and Random Forest(RF).A hyperparameter tuning method based on Sequential Model Global Optimization(SMBO)is proposed to obtain the optimal hyperparameters of the lane-changing decision model,which solves the problem that the machine learning algorithm tuning process is complex and cumbersome,and the global optimality of the parameters cannot be guaranteed.The results show that,in terms of accuracy,receiver operating characteristics(ROC)and other indicators,the lane-changing decision model based on LSTM has excellent performance,and the accuracy rate on the test set can reach 97.71%.The performance of the GRUbased lane-changing decision model is similar to the LSTM-based lane change decision model,and the accuracy on the test set can reach 97.64%,which is better than the existing algorithms used for comparison.(3)Considering the applicability and versatility of the lane-changing trajectory planning algorithm in different road shapes,we study the conversion method between the Frenet coordinate system and the Cartesian coordinate system,and proposes a lanechanging trajectory planning method based on quintic polynomial.Using the deep learning method to predict the movement state of surrounding vehicles,consider the future traffic state for lane-changing trajectory planning,and based on the Gipps collision avoidance model,put forward the end safety zone of the lane-changing trajectory planning,thereby avoiding the potential collisions during the lane-changing process,enhances the safety of the lane-changing process.Based on the safety zone at the end of the lane change trajectory and the quintic polynomial trajectory planning method,a multiobjective optimization lane change trajectory planning method with safety,comfort and efficiency as the constraints and optimization goals is proposed,and the interior point method is used to solve the lane changing duration time.(4)Design Stanley controller and PID controller for path and vehicle speed tracking control,build a Car Sim and Simulink Co-simulation experiment platform,and design two bend lane-changing scenarios under normal and congested traffic flow based on real traffic flow data to verify the performance of the proposed lane-changing trajectory planning and trajectory tracking control algorithm.The simulation results show that the proposed lane-changing trajectory planning and tracking control algorithm can successfully complete lane-changing under normal and congested traffic flows,and the prediction-based lane-changing trajectory planning method can maintain a greater distance to the target lane following vehicle TR during the lane-changing process.The minimum distance to the vehicle TR in the normal traffic flow lane-changing scenario is increased from 3.84 m to 15.53 m,and the minimum distance to the vehicle TR in the congested traffic flow lane-changing scenario is increased from 7.16 m to 7.92 m.