Research On Humanlike Decision Making And Motion Planning For Lane Changing For Multi-Objective Scenarios

Autonomous vehicles(AVs)can reduce traffic accidents and improve transport efficiency,which have been highly concerned by governments.However,due to the technology maturity and huge costs,AVs and human driving vehicles will coexist in the traffic environment for a long time in the future.At this point,the non-humanlike lane changing(LC)behavior of AVs may mislead human driving vehicles.Thus,AVs may cause traffic flow instability and even cause traffic accidents.In addition,the driving behavior that only pays attention to physical safety ignores the cognitive safety which is limited by the physiological and psychological endurance of passengers.As a result,the passengers’ acceptance of AV is low.To solve this problem,numerical optimization algorithms with high-dimensional and nonlinear or learning algorithms are used to approximate human driving behavior with a poor real-time performance on the limited on-board computing resources.On the contrary,low-dimensional and linear algorithms have poor personification ability.Because the essence of artificial potential field(APF)method is to represent that people are subjected to virtual social forces,it has high real-time and strong humanlike characteristics in theory.Thus,it has become a potential method to solve the above problems.Based on the APF,the following research is carried out in four aspects: the design of potential field function,the construction of driver’s subjective LC intention mechanism,the virtual lane lines motion planning algorithm and the experimental verification.Comfort is an important aspect of humanlike behavior.However,the control acceleration generated directly by the potential field gradient is always too large,which deteriorates the cognitive safety and comfort of passengers.In order to accurately reveal the high-dimensional influencing factors and coupling relationship of potential field,firstly,principal component analysis and correlation analysis are hired to calculate the main influence factors and variable orders with the naturalistic driving data(NDD).Therefore,a virtual spring damping action model is proposed to represent the car following behavior and lane keeping behavior.The model can accurately quantify the driver’s cognitive risk,speed requirement and has the advantage of parameter interpretation.Finally,in order to accurately describe the coupling relationship between dynamic car following distance and multiple factors in the spring damping force model,two methods are designed to improve the calculation accuracy of car following distance in dynamic scenarios.The one is a steady-state car following distance based on Gaussian mixture model(GMM).The other one is a dynamic correction technology considering safety and efficiency.The virtual force basis function of potential field lays a foundation for the research of the following chapters.At the decision-making level,the potential field value is used as the safety index to design the lane change intention,lacking the influence of driving efficiency and multi-target objects,which is mainly applicable to the few target scenarios.Firstly,based on the LC data,the significance analysis is used to reveal that the driver’s lane change decision has the characteristics of considering both safety and speed requirements.Secondly,combining this characteristic with the virtual force basis function of the potential field,a conversion coefficient is designed to linearly map the longitudinal potential field force to the lateral driver tolerance force(DTF),and a lane change decision-making strategy based on the DTF is proposed.Finally,by analyzing the distribution characteristics of different objects in the visual area,a visual attenuation coefficient is designed to modify the DTF generated by different objects.It achieves both LC safety and speed requirements while ensuring the accuracy of LC intention in multi-objective scenariosAt the planning level,considering that the lateral potential field force of LC is unknown and its action mode is unknown.Thus,LC motion planning cannot be directly realized.Firstly,the experiment of the relationship between driver’s LC behavior and visual focus is designed.And the characteristics of driver’s visual focus shift are verified by using correlation analysis method.Secondly,according to this characteristic and the force basis function of the potential field,a virtual lane line model representing the driver’s lateral expected safety boundary is constructed.Based on theoretical derivation of its movement form,a LC planning algorithm based on virtual lane lines is proposed.Finally,the significance analysis method is used to determine the main influencing factors of the virtual lane line movement characteristics.And a LC duration estimation method based on the multi parameter bounded GMM is proposed.The analytic solutions of lane change trajectory,velocity and acceleration are obtained,which can ensure the real-time performance of the algorithm and consider the safety and comfort of motion planning.A real vehicle and simulation experiment platform are built to verify the effectiveness of the humanlike LC decision and planning algorithm proposed in this paper from the perspective of safety,comfort and efficiency,considering low-speed and high-speed scenarios.The experimental results show that the algorithm has good performance in humanlike LC decision-making and planning in different vehicle speeds and multi object scenarios.