Research On Reasoning Decision-making And Self-optimizing Control Algorithm Of Intelligent Driving Bus In Lane Changing Scenarios

Autonomous driving vehicle is an electrified and automatic mobile platform integrating intelligent perception,autonomous decision-making,rapid planning and optimized control.It has become an important technology development direction for enterprises and scientific research institutions.As a typical autonomous vehicle,intelligent bus runs in dynamic and random complex traffic scenes,and uncertainty factors such as sensor measurement error and internal and external disturbance of the vehicle,which lead to great challenges on the safety and comfort driving of intelligent bus.This dissertation focuses on the behavior decision and optimal control of autonomous driving buses in lane-changing scenarios.In order to improve the robustness of decision making and the adaptive ability of control,we present a security situation valuation model based on convolution neural network,a lane change decision model based on bayesian reasoning and prior knowledge,and a self-optimizing motion controller based on reinforcement learning.Firstly,we establish the hardware-in-the-loop simulation platform and intelligent driving bus platform.According to the people-vehicle-road-environment four dimensions,we expand five typical urban road categories,covering 1000 traffic scenarios.The perception-decision-control system module can be tested in the whole process of closed-loop,which can quickly verify the safety of the algorithm in random lane change scenarios.We have built an intelligent driving bus with a length of 8m,equipped with high-precision positioning combined inertial navigation(INS),long and short range detection lidar,360° circular camera and other sensing devices and low-power onboard hardware computing unit.In addition,intelligent driving software system and data communication module were developed to enable the vehicle to have automatic driving capability,which is used to verify the effectiveness and driving comfort of safe lane change in real road scenarios.Secondly,in order to poercive the distance and predict the speed of surrounding traffic vehicles in the process of lane changing and obtain the best lane changing time,we proposed a safety assessment model based on convolutional neural network,which comprehensively considered the attributes of typical urban road traffic scenes and realized the feature description and safety evaluation of driving situation.On the one hand,the convolutional network based multi-label task training mechanism and the gradient descent algorithm are used to complete the mining of high-level semantic feature information in the image,so as to predict the longitudinal relative distance between the traffic vehicle and the host vehicle.On the other hand,a hybrid haussian observer was proposed to characterize the speed distribution characteristics of traffic flow,and the maximum likelihood estimation algorithm was used to complete the parameter estimation of multivariate Gaussian model.By constructing the safety threshold evaluation function,the safety degree under the current driving situation was evaluated.The simulation results show that the average absolute error of longitudinal distance prediction of obstacles in front of vehicles is less than 3m within30 m distance,and the actual vehicle test results show it is less than 2.5m in the range of 50 m.Meanwhile,the distance prediction results can be dynamically evaluated after inputting the safety threshold function,and the objective analysis results show that the safety evaluation is consistent with human drivers.Thirdly,in order to overcome the uncertain influence of the prediction error of visual perception system on the decision result and further improve the safety of lane change decision,this dissertation proposed a Bayesian decision agent model(BDA)based on prior knowledge and real-time data,inspired by the cognitive decision rules of human drivers.The posteriori probability distribution of lane change decision is inferenced by probability propagation algorithm,and realized the robust decision of intelligent driving bus in complex scenes.Specifically,we establish a probabilistic grid map for typical three-lane changing scenarios,and propose a heuristic search algorithm(KB-GES)to search the optimal probabilistic reasoning model from the sample data of drivers’ lane-changing behavior.The similarity between the agent’s decision model and human driver’s decision is verified by the intragroup correlation coefficient(ICC).Experimental results of typical lane-changing scenarios show that the correlation coefficient ICC between BDA model and human driver reaches 0.984,which reduces the dependence on the prediction accuracy of visual sensor,and deduces relatively robust driving decision mode under the constraints of safety and driving efficiency.Finally,in order to overcome the uncertainty caused by internal and external disturbances and reduce the influence of steering wheel fluctuation on vehicle comfort during lane change movement,we introduce the environment interactive self-learning mechanism into the vehicle control system,and present a self-optimizing controller based on reinforcement learning,which continuously enhanced the optimization ability of behavior control during the interaction between the vehicle and the environment.Specifically,we exploit the improved deterministic strategy gradient algorithm to train Actor-Critic network,and integrate the classical PID controller constraint to solve the balance problem of exploration and utilization.By calculating the gain value of PID control parameters,the intelligent bus can compensate the control deviation in moving process,and then improve the adaptive ability of the controller to disturbance.In the real vehicle experiment,the driver behavior data cache area is created to construct the positive sample state space,and the state switching mechanism is adopted to ensure the stability of the reinforcement learning self-optimizing controller.The experimental results show that the controller can optimize the control parameters in real time.In the simulation test,the standard deviation of the lateral position of the self-optimizing motion controller based on reinforcement learning is 0.071 m,and in the real vehicle test,the standard deviation of the lateral position is 0.272 m.The test results of steering wheel fluctuation standard deviation in simulation and real vehicle are 0.04° and 80.69° respectively.