Research On Lane-changing Trajectory Planning Strategy For Automated Driving Considering The Interaction Between Ego Vehicle And Traffic Vehicles Behavior

High-level automated driving vehicles represent the trend and forefront of automobile industry development,and represent the strategic high ground of future automobile technology.Trajectory planning is important basis for promoting the application of high-level automated driving vehicles,which can effectively ensure the safety,efficiency,and comfort of automated driving vehicles,and have become important research topics in related fields.In recent years,with the rapid development of artificial intelligence technology,trajectory prediction of traffic vehicles and trajectory planning of autonomous vehicles have shifted from traditional mechanism model methods to data-driven methods,achieving significant progress in single vehicle intelligence.However,in the dynamic and complex traffic environment with strong random uncertainty,existing technologies still face challenges in achieving human-like interactions with other traffic participants.Therefore,how to endow autonomous driving vehicles with the ability to interact and drive like humans,in order to improve user adaptation,acceptance,and trust,and reduce safety hazards caused by human-machine differences,has become a core scientific problem urgently to be solved in the field of trajectory planning and trajectory prediction.This paper analyzes the driving prediction and decision-making mechanism of skilled human drivers when changing lanes,proposes an autonomous driving lane-changing trajectory planning strategy that considers the interaction between the ego vehicle and traffic vehicles behavior,takes the lane-changing scenario as the main research scene.The validation was conducted on a multi-driver-in-the-loop simulation platform in a self-built Windows-Linux cross-platform environment based on VI-grade hardware and software:The main research contents of this paper are as follows:(1)Data collection and processing of lane-changing interaction based on multi-driver simulation platform.First,by comparing the data collection methods commonly used in the prediction and planning fields with research needs analysis,a technical route of using a driving simulator for data collection was determined.Next,to solve the problem of lack of realism of the simulator,a high-precision vehicle dynamics model was established based on an SUV prototype,providing a basis for achieving a realistic driving experience.Then,a multi-driver simulation platform based on VI-grade software and hardware was built,which improved the driving immersion from various aspects such as visual,auditory,tactile,and interaction.Finally,a data collection experiment was designed and the data collection and cleaning was completed,generating the final dataset for subsequent model training.(2)Design of basic trajectory prediction and planning strategy based on vector codingA multi-modal basic trajectory predictor model and a basic trajectory planner model based on lane sequence planning were designed and trained on the self-developed dataset.First,the hierarchical graph neural network Vector Net was introduced for vectorizing traffic elements,and the process of subgraph and global graph feature extraction was described in detail,explaining how the encoded environmental context information was used for subsequent basic trajectory predictors and planners.Then,the principle and implementation of the multi-modal trajectory predictor were explained in stages,and a probability model was established by transforming future trajectory prediction into a distribution problem of predicting discrete target endpoints,which were used as conditions.Finally,a basic trajectory planner based on lane sequence planning was proposed.The planner consists of three end-toend networks: a lane sequence planning model,a motion distance planning model,and a human-like trajectory generation model,which can output planned-level vehicle trajectories.(3)Trajectory planning strategy considering active interaction between the ego vehicle and traffic vehiclesFirst,the interaction problem between the ego vehicle and the traffic vehicles is modeled,and the joint distribution is decoupled into the marginal distribution of one traffic participant and the conditional distribution of the other traffic participant.Then,by analyzing the human driving mechanism,it is found that the driver makes different assumptions about the behavior of other vehicles and flexibly considers their own driving behavior from multiple hypotheses.Therefore,a proactive interaction relationship discrimination rule is designed,and an encoderdecoder architecture proactive interaction relationship discrimination model is trained,achieving an accuracy of 91%.Furthermore,a traffic trajectory prediction and ego vehicle trajectory planning interactive strategy is designed,and the encoder is modified to enhance the scene-encoding the future trajectory of the proactive vehicle into the feature of the passive vehicle based on the discrimination result,so as to learn the interaction behavior performance in the human driving mechanism and achieve more reasonable trajectory prediction and planning.(4)Validation of lane-changing trajectory planning strategy considering the interaction between the ego vehicle and traffic vehiclesFirstly,the metrics of the interaction strategy are introduced and appropriately improved.Then,algorithmic models based on datasets were verified.In order to verify the predictor in the interaction strategy,four publicly open prediction datasets were compared.Considering the research requirements,Argoverse1.0 was finally selected as the dataset for comparison experiments in this chapter,and was compared with classical trajectory prediction algorithms,proving the advanced performance of the interaction predictor.The effectiveness of the interaction strategy was demonstrated through ablation experiments.Compared with the traffic vehicle prediction and trajectory planning without considering interaction,the interaction strategy had better accuracy and improved the rationality of the results.Finally,to illustrate the effectiveness of this research in practical applications,a driver-in-the-loop simulation was conducted in a Windows-Linux cross-platform environment based on VI-grade hardware and software,simulating real-world driving scenarios.The experiment proved that this strategy can effectively plan human-like lane-changing trajectories in interactive scenarios,ensuring vehicle safety and driving efficiency in heterogeneous,complex,and highly dynamic driving environments.