| The development of intelligent vehicles is one of the important ways to solve traffic safety and efficiency,and high-level autonomous driving is the high point of strategic competition among the world’s industrial powers.As the top and bottom part of the whole system,the intelligent or not of the autonomous driving decision algorithm largely determines whether the autonomous driving system can be used or not,and also largely determines the driver’s acceptance of the vehicle intelligent system.Traditional decision-making algorithms are mostly based on rules and machine learning methods,with the following difficulties:(1)too much reliance on large-scale trajectory data sets,poor utilization of the sensors on the car itself,and poor verifiability of the algorithm;(2)in the face of complex,new road scenarios often do not learn and make decisions well,and the applicability of the algorithm is poor;(3)at the beginning of the algorithm design does not take into account the different(3)the personalized needs of drivers for vehicle decision making are not considered at the beginning of the algorithm design,and the decision results are too aggressive or conservative,which affects the trust of drivers in the vehicle decision system.Therefore,it is difficult for the traditional self-driving decisionmaking algorithm to meet the self-driving needs of the rapidly developing automobile industry.In this paper,based on the "National Natural Science Foundation of China Joint Fund Key Project",we have conducted an in-depth research on the lane change decision method of intelligent vehicles considering driving style:(1)A driver decision dataset integrating real scenes and virtual environments is constructed,enriching the data content and dimensionality;a statistical model of macro and micro driving behavior features with dual temporal windows is innovatively applied,and a multimodal spatiotemporal structure driving style feature learning method based on convolutional long and short time networks is designed,solving the problem of insufficient utilization of temporal features and achieving a 93.81% accuracy in the independently constructed dataset.accuracy,which is a16.61% improvement compared to a single convolutional network,and the experimental results show an increase in utilization for each data type and still strong robustness in real car data.(2)Combined with the content of the dataset,the characteristics of driver driving data under different driving styles were analyzed,and the lane change decision feature parameters were selected to consider the vehicle lane change decision process from four dimensions of travel efficiency,ride comfort,speed gain and driving safety,and the weight coefficients under different driving styles were set,and the input was made stylized by coefficient adjustment,based on the multimodal long and short term memory network The LSTM-DLC lane change decision model is established based on the multimodal timing characteristics of the long and short term memory network,and the driving style characteristics are introduced.The experimental results show that the introduction of the driving style improves the accuracy of the decision by 8.53% and has a large application value.(3)Real-world verification of the driving style recognition and lane change decision algorithm was completed.Based on the ROS platform of the intelligent driving platform "Jiangda Smart Line II",the two algorithms proposed in this paper were deployed,and the algorithms were verified under real road scenarios.The proposed model can identify the driver’s driving style type more accurately and make anthropomorphic lane change decision under real scenarios. |
PDF Full Text Request