Research On Interactive Decision Making And Trajectory Planning For Intelligent Vehicles Considering The Driving Characteristics Of Traffic Participants

Intelligent driving technology is revolutionizing the transportation system in ways never seen before.Including core elements such as perception and positioning,decision making and trajectory planning,and tracking control,among which decision making and trajectory planning serve as the“brain” of intelligent vehicles.Its goal is to select the best driving strategy based on sensor information through intelligent decision making and trajectory planning,ensuring the vehicle’s safe and efficient arrival at its destination.However,intelligent vehicles are not the only road users.Only when intelligent vehicles have the capability to interact with various dynamically changing traffic participants can they collaborate and share the road safely and adaptively with surrounding traffic participants.The decision making and trajectory planning system aims to achieve target lane selection and desired reference trajectory generation,and is a key technology for ensuring safe,efficient,and traffic-rule-compliant vehicle interactions,hence,it is necessary to research into interactive decision making and trajectory planning methods.Thus,this paper takes the intelligent vehicle decision making and trajectory planning system as the research object,focusing on the mining and utilization of inter-vehicle interaction information in the prediction of trajectories for traffic participants,intelligent vehicle behavior decision-making,and trajectory planning systems,exploring traffic participant trajectory prediction,intelligent vehicle behavior decision under single-vehicle perception,intelligent vehicle behavior decision under network connected conditions,and intelligent vehicle trajectory planning.The specific research content includes:Firstly,to address the challenges in recognizing the dynamic random characteristics of traffic participant trajectories and the limited accuracy of trajectory predictions due to complex and variable inter-vehicle relationships,a multi-model algorithm based on an adaptive state transition probability matrix is proposed.Real traffic participants’ data collection experiments are organized,statistical methods are used to analyze and select features that effectively represent driving characteristics,a set of motion models for vehicles with different driving modes are built.Based on the collected data,gap satisfaction probability function and driving mode probability function are designed,the two functions are used to adaptively update the state transition probability matrix,which is then integrated into the multi-model algorithm.This method can obtain the driving mode probability of traffic participants based on the match probability between the actual motion state of traffic participants and the preset model,achieving probabilistic representation of traffic participant behavior recognition results and personalized trajectory prediction,the behavior recognition delay time is deduced,and the trajectory prediction accuracy is greatly improved.Secondly,to address the issues of insufficient consideration of various driving characteristics of traffic participants and poor adaptability of decision results under single-vehicle perception,a layered game decision-making method considering the politeness level of traffic participants is proposed.Driving scenarios are allocated with multiple leader-follower identities according to spatial relationships and influence mechanisms.Acceleration estimates of surrounding traffic participants are obtained using Kalman filtering,and the probability of their politeness are derived based on the spatio-temporal relationship with the vehicle and acceleration/deceleration behavior under that relationship.The recognition results of politeness levels are used to adaptively adjust the weights of safety,comfort,and efficiency in the cost function of traffic participants.Moreover,the convexity of the designed traffic participant cost estimation function is constructed and proven based on parametric selection.Intelligent vehicle infers their optimal behavior based on the traffic participant cost estimation function and adaptively adjust its behavior decisions,achieving adaptive interactive behavior decision making that integrates the politeness of traffic participants.Thirdly,to address the challenges of high-dimensional traffic information fusion under network connected conditions and the inability to evaluate the reliability of decision-making results,a cognitive uncertainty-aware intelligent vehicle behavior decision method under network connected conditions is proposed.The original sample set is divided into multiple subsets using the bootstrap resampling method,resulting in an ensemble of deep neural networks trained to form a value estimation distribution for decision results.To avoid the homogenization tendency in value estimation by multiple ensemble neural network members,a random prior network with fixed parameters is integrated into the neural network structure initially,enhancing the independence and heterogeneity of multiple ensemble neural network members in estimating the value of decision-making results,ensuring the effectiveness of uncertainty assessment of available decisions.The relative standard deviation of multiple estimated values is introduced to construct the decision mode switching condition,prompting intelligent vehicles to switch decision-making modes in scenarios of insufficient exploration,ensuring the reliability and safety of decision results.Finally,to address the high computational complexity and time consumption of threedimensional(X-Y-T)trajectory planning,an adaptive sampling-based two-layer three-stage trajectory planning method is proposed.The three-dimensional trajectory planning problem is decomposed into two-dimensional optimization problems of path planning and speed planning based on the Frenet coordinate system.In path planning,an adaptive sampling method based on artificial potential field is used to reduce sampling points in the longitudinal and lateral position space,and the path with lowest cost is calculated using dynamic programming,further optimizing this path by constructing a quadratic programming problem.In speed planning,adaptive sampling using kinematic constraint is performed,and the position-time curve with lowest cost is determined using dynamic programming,constructing a quadratic programming problem based on the predicted trajectories of traffic participants and kinematic constraints to obtain the final speed curve.Finally,based on a two-degree-of-freedom vehicle dynamics model,a trajectory tracking error model is built,and lateral and longitudinal controllers are designed to verify the actual tracking performance of the planned trajectory in the tracking control level.Simulation results show that the proposed trajectory planning method can effectively reduce the search space and shorten the time cost,balancing the quality of the planned trajectory and the speed of the solution.After verifying the feasibility of the overall algorithm on a joint simulation platform,the hardwarein-the-loop experiment involving real human driver as traffic participants demonstrated that the trajectory prediction,decision-making,trajectory planning,and tracking control strategies proposed in this paper can achieve integrated collaboration.