Research On Personalized Driver Assistance Algorithm For Intelligent Vehicle

Intelligent vehicle is a complex system that includes some subsystems,and has the characteristics of strong coupling and conflict among these subsystems.With the development of automobile automation,there is an increasing trend that more and more driver assistance technologies are employed in automobiles.The interactions between human driver and intelligent driving systems and vehicle electronic control system are becoming increasingly complex.In the process of achieving the fully autonomous driving vehicles,the co-existance of human driver and intelligent driving technologies will last a long time.Drivers' driving characteristics are important for the design of intelligent driving technologies.Embedding drivers' diverse characteristics into a traditional assiatance system and making the controlled vehicle natural to human drivers are of vital importance to a comfortable and enjoyable driving experience.Because of the large size of human driver population,the difference in drivers' character,psychological state,driving skill and the inherent complexity and randomicity in human driving behavi or,it is difficult to reflect drivers' characteristics in traditional assistance systems by calibrating the control parameters.With the fund of the National Key Research and Development Program of Chi na-“Theory of Human-Machine Interaction in Intelligent Electric Vehicle”(No.2016YFB0100904),the National Natural Science Foundation of Chi na-“Research on Human-Vehicle Conflict and Cooperative Control for Intelligent Electric Vehicle”(No.51775235),this paper proposed the learning-based personalized driver assistance strategy.Based on human driving data,the characteristics of human driving are studied.And learning-based methods are used to design the upper driver assistance algorithms where drivers' driving characteristics are taken into consideration.The coordination of upper longitudinal/lateral assistance decisions and vehicle dynamics control is taken as a multi-objective optimization under some constraints.A numerical optimization algorithm is established to achieve the control allocation.And both targets of driver assistance systems and vehicle electronic control are realized.This paper mainly focuses on the key problems of personalized driver assistance strategies.The research is carried out mainly from the following aspects: the driving data collection and similarity measurement of different driving styles,longitudinal driver assistance method that can consider the drivers' styles,personalized lateral driver assistance strategy and vehicle chassis integral control method.Firstly,a novel method is proposed to measure the similarity of different driving styles from the driving data distribution characteristics.Driving style analysis is a fundamental problem in the design of personalized driver assistance strategy.Traditional driving style clustering method measure the difference of different human drivers based on the statistical characteristics of vehicle speed,acceleration and steering angle.The Euclidean distance of statistical characteristics is taken describe the similarity of driving styles.However,the randomness of driving behavior cannot be considered in this method.The statistical characteristics,such as mean value,maximum and minimum values,cannot describe the driving characteristics of driving behavior.The driving data can't be fully applied.Concerning this problem,the data-based analysis method is utilized by taking the driving data of every driver as a probability distribution function,fitted with a Gaussian Mixture Model(GMM),which are learned using expectation-maximization(EM),an efficient iterative method.The Kullback-Leibler(KL)divergence is applied to compute the difference of different GMMs.Then,a novel KL divergence-based unsupervised clustering algorithm is proposed to cluster human drivers into different categories,where the similarity of drivers is indexed by the KL divergence.The problem of driver's lane-changing intention identification is studied in this paper,where the nonlinear relations between vehicle states and driver's lane-changing intention are analyzed by applying mutual information theory.And a hidden Markov model is established by learning from human driving data to estimate the lane-changing intention.Secondly,a longitudinal assistance strategy is proposed based on adaptive dynamic programming.Where drivers' driving characteristics and the randomness of preceding vehicle can be taken into consideration.The control of longitudinal assistance strategy is divided into two control modes: speed control and car-following control.An exponential speed control method is proposed in the speed control mode,which can achieve the acceleration performance similar to that of human drivers.In the car-following control mode,a mapping model between vehicle states and driver's longitudinal decision is established.The inverse reinforcement learning method is utilized to learning the desired vehicle acceleration of human driver in the car-following scenario.It is assumed that drivers' longitudinal decision satisfies the Boltzmann distribution,and the maximum likelihood inverse reinforcement learning is used to learn the decision of human drivers in car-following condition.To model the longitudinal driving environment,an undirected probability graph model is established based on the human driving data.The decision of vehicle longitudinal control is realized by using adaptive dynamic programming.An actor module and a critic module are established using artificial neural network.Through interactive iteration between car-following control algorithm and the traffic model,a learning-based control method is established in this paper.Thirdly,a personalized lateral assistance algorithm is proposed based on vehicle trajectory prediction and deep reinforcement learning.An integrated framework to address the problems of driving-intention estimation and vehicle trajectory prediction is proposed based on DBN,where the unobservable lane-changing intention variable,driving behavior and vehicle states are semantically included.An SIR particle filter is used,including two nested loops,to infer vehicle trajectory from the DBN.Deep learning technique has the outstanding performance in feature extraction,and has achieved state-of-the-art results in some research fields.Deep convolution neural network is employed in this paper to establish a driver model by learning from the driving data in the platform of CARLA.Deep reinforcement learning is used to design the lane keeping strategy by interacting with the driver model in CARLA.In order to improve the efficiency of deep reinforcement learning,the variational autoencoder is used to extract the features from image.Finally,the developed lateral assistance algorithm is verified in CARLA.Fourthly,a chassis integrated control strategy is proposed based on multi-objective optimization technique to coordinate the upper driver assistance strategy and vehicle electronic control system and avoid the possible conflicts of intelligent driving strategy and vehicle stability.The reference vehicle control in a future time horizon is predicted using recurrent neural network.And the control allocation is achieved using numerical computation considering the vehicle dynamic model and the actuators' dynamic characteristics.Simulation results show that good performance and computation efficiency can be obtained.In addition,a driver-in-the-loop platform is established using CARLA,Carsim and the hardware platform of d SPACE/ Simulator,Micro Auto Box.Some experiments are conducted to verify the proposed personalized driver assistance strategy.The driver-in-the-loop platform is useful in intelligent driving technique test with the functions of vehicle dynamics simulation and 3D scenario construction.Based on the driver-in-the-loop platform,some exprements are carried out in different scenarios.Results show that both longitudinal and lateral control target can be reached.A real vehicle experiment platform is established to test the longitudinal assistance strategy.In summary,the main innovations of this paper are as follows:(1)A driving similarity measuring method is proposed based on the characteristics of driving data distribution.Different from traditional driving style studies that distinguish human drivers based on statistical metrics,human driving data is taken as a probability distribution function,fitted with GMM,and the difference between GMMs is introduced to measure the similarity of human drivers,which provides an effective route to quantitatively interpret the similarity of different driving styles.(2)A personalized longitudinal driver assistance strategy is proposed based on human driving style.The mode switching rule is established based on driver action points extracted from real-world driving data.The speed control algorithm is studied based on human driving data.The design of car-following strategy is taken as sequential decision problem in dynamic traffic environment considering the driving characteristics of human driving.The learning of car-following control is modeled as a multi-objective decision problem,such tracking,ride comfort and driving style adaptability based on the theory of adaptive dynamic programming.The characteristics of human driver in longitudinal driving is studied based on human driving data using the inverse reinforcement learning.The randomness of preceding vehicle in modeled as a probability graph model.The car-following algorithm is established by interacting with the traffic model.(3)A personalized lateral driving assistance strategy is proposed based on ve hicle trajectory prediction and deep reinforcement learning.The relation of some variables are modeled as a dynamic Bayesian network that can be used to achieve an accurate prediction performance,which is import to the decision of lateral driver assistance strategy.A method for designing personalized driver assistance strategy is proposed.The characteristics of human driving is learned using deep neural network,and the assistance strategy is achieved by interacting with the driver model.