Lane Change Risk Assessment And Decision Method Based On Driving Style Identification And Motion Prediction

The expressway has the characteristics of faster driving speed and easy collision in the process of following the car and freely changing lanes.At present,the traditional lane change decision model in the highway scene only considers the completion of the driving task,which is not fully in line with the decision mechanism adopted by human drivers in driving behavior decision that balances safety and traffic efficiency,and fails to fully consider the surroundings uncertainty of vehicle behavior brings collision risk to the host vehicle.Therefore,in order for autonomous vehicles to operate safely and reliably in dynamic uncertain scenes on highways,autonomous driving systems need to have the ability to predict the movement of other traffic participants and assess the potential risks posed by their behavior.This paper mainly studies the risk assessment and behavioral decision-making of autonomous driving vehicles considering the driver style and long-term motion prediction of surrounding vehicles in the highway scene.The research focuses on four aspects:analysis of driving style of human-driven vehicles,intent recognition and probabilistic trajectory prediction of other vehicles in the scene,collision risk assessment and decision-making of autonomous driving behavior.Firstly,the driver's style characteristics are extracted from the historical trajectory of the vehicle,and the cluster analysis of the driver's style of the human-driven vehicle is performed.Compare the clustering effect of K-Means++,fuzzy mean(FCM)and hierarchical clustering algorithm under multiple clustering centers,Use support vector machine(SVM)to classify driving style.Analyze the operating characteristics of the three styles of drivers when changing lanes and following the car.The results show that using the K-Means++ algorithm to cluster test datasets of the three driving styles is greater than 97%.Drivers of different styles have significant differences in the execution time of lane change,the lateral and vertical acceleration during the lane change,and the headway distance in the following state.The driver with the more aggressive driving style has a higher frequency of lane change,a lower lane change time and the speed stability is lower during the lane change,and the cautious driver has a higher headway when following the car.Secondly,extract the historical trajectory features of vehicles from the NGSIM dataset,identify the intention of human-driven vehicles to change lanes,and make probabilistic predictions of driving trajectories.Bidirectional Long Short-Term Memory(Bi-LSTM)neural network is used to model driving intentions,and two input feature forms are compared,including only target vehicle information and historical trajectory information of surrounding vehicles.The results show that when the prediction duration is 3.5 seconds,the latter's intention recognition accuracy rate is 10.02%higher than the former.Intentional recognition was performed on different recognition durations,and it was found that the recognition accuracy rate gradually decreased as the prediction duration became longer.Among the models incorporating historical trajectory information of surrounding vehicles,the accuracy of the model with a prediction time of 3.5 seconds is reduced by 23.59%compared to the model with a prediction time of 1 second.LSTM combined with Mixed Density Network(MDN)model is used to predict the distribution of vehicle driving trajectories.The prediction results of the model are compared with the results of other related work,which shows the effectiveness of the model in predicting vehicle motion.The traffic flow video is collected from the Jinhua South Road section of the second ring road in Xi'an,and the input features of the motion prediction model are extracted.Then the features are input into the trained model,and the generalization performance of the above motion prediction model is verified.Finally,a risk assessment and lane change decision model is established.Trained the target lane selection model of autonomous vehicles.Proposed a vehicle trajectory conflict quantification method based on collision time(TTC)and headway time(TH).The conflict between the surrounding vehicle trajectory obtained by the trajectory prediction model and the trajectories corresponding to the different intentions of the host vehicle is quantified,and the conflict risk of the different trajectories of the host vehicle is obtained;By screening the candidate trajectory cluster of the host vehicle,make driving decisions.The feasibility of the risk quantification and decision model is verified in the decision scenario of the dataset.