| During the human-machine cooperative driving phase,the driver and the automatic driving system should be in a mode of coordinated communication and complementary advantages.However,the premise of realizing such a mode is that the automatic driving system needs to obtain the state data of the driver’s intention and operation in time,otherwise various types of conflicts may occur,which may cause forced switching control or even cause an accident.In this paper,the lane change phase in the human-machine cooperative driving mode is taken as the research object.When the driver starts to perform the lane change operation,the automatic driving system should acquire the driver’s current lane change execution mode in a timely and accurate manner.Common lane change execution modes include: forced lane change with a following vehicle in the target lane,unsuccessful lane change behavior,waiting and changing lanes till the following vehicle in the target lane passed.From the perspective of pattern recognition,it is necessary to recognize whether or not the lane change operation is performed,so it is necessary to recognize the lane keeping behavior.In this case,how to realize the fast and accurate recognition of these modes by the automatic driving system is an important issue that affects the safety of the lane change behavior in the human-machine cooperative driving mode.In response to the above requirements,this paper built a multi-sensor test acquisition platform,which integrates AWS(Advanced Warning System),IMU02 gyroscope,VBOX-3I differential GPS,millimeter wave radar and some other sensors.This paper recruited 43 drivers to conduct experiments on urban low-speed trunk roads,urban expressways and highways.980 sets of four types of driving behavior process sample data were selected,and the durations and pressing line times of the three types of driving behaviors except the lane keeping behavior were statistically analyzed.Kalman filtering was performed on the selected sample data,and the characterization parameters of the four types of driving behaviors were analyzed.The filtered characterization parameters were normalized and analyzed with a principal component method to provide better basic data for subsequent Support Vector Machine(SVM)modeling.The main research contents and conclusions of this paper are as follows:1.By using the distance between the subject vehicle and the left lane line,the steering angle,the relative distance between the subject vehicle and the following vehicle in the target lane,and the distance between the subject vehicle and the left lane line,the steering angle,the relative distance between the subject vehicle and the following vehicle in the target lane,the relative speed between the subject vehicle and the following vehicle in the target lane as inputs respectively,the driver lane change execution pattern recognition model based on the decision tree classifier is established.However,the overall recognition accuracy of the model is low and the real-time performance is poor.The highest overall recognition accuracy of the four types of driving behaviors is 77.55% of the 1.8s time window under the three-parameter combination.After pruning,the effect is even worse.2.For the decision tree model,the effect is poor.The random forest classifier is used to establish the driver lane change execution pattern recognition model.Similar to the decision tree,the three-parameter combination and the four-parameter combination are still as inputs,and the overall recognition accuracies of the four types of driving behaviors in each time window are compared.The three-parameter combination recognition is better than the four-parameter combination,and the degree of superiority is better than the decision tree.That is to say,the relative distance has a certain degree of representation coverage for the information contained in the relative speed,and the addition of the relative speed will make the recognition effect of the model worse.The highest overall recognition accuracy of the four types of driving behaviors of the three-parameter combination is 3.58% higher than the highest overall recognition accuracy of the four types of driving behaviors of the four-parameter combination,and the corresponding recognition time window reduces by 0.7s.In the 1.2s time window under the three-parameter combination,the overall recognition accuracy of the random forest model for the four types of driving behaviors is the highest overall recognition accuracy and its value is 88.78%.The real-time performance is good,and the recognition effect is better.3.Considering that many researchers at home and abroad use SVM to recognize the lane change behavior,in order to explore the advantages and disadvantages of the recognition model based on the random forest classifier,this paper establishes the driver lane change execution pattern recognition model based on SVM that the grid search method and genetic algorithm are used to optimize respectively.Comparing the SVM model with the random forest model,this paper obtains the driver lane change execution pattern recognition model based on the random forest classifier is a better choice,and its recognition effect is much better than SVM.For the highest overall recognition accuracies,the difference between the random forest model and SVM with the grid search method and genetic algorithm optimizing are 20.92% and 20.41%,respectively.4.In order to further explore the classification performance of SVM,the grid search method and genetic algorithm are respectively used to optimize SVM,and the three types of driving behaviors and two types of driving behaviors are recognized by SVM.Comparing the recognition results of SVM for three and two types of driving behaviors with the recognition result of SVM for four types of driving behaviors,it is found that when the classification category of SVM is reduced,whether it is the grid search method or the genetic algorithm for optimization,the SVM classification ability is improved significantly.The average values of the differences between the overall recognition accuracies of the two types of driving behaviors and the overall recognition accuracies of the four types of driving behaviors are analyzed.The improvements of the overall recognition accuracies of SVM under the two optimization methods are more than 15% whether it is a three-parameter combination or a four-parameter combination. |
PDF Full Text Request