Research On Path Tracking Control Algorithm For Autonomous Ground Vehicle

Autonomous ground vehicle(AGV)technology has developed rapidly in recent years and has become a hot research direction at home and abroad.There are two main designs for AGV.One is a rule-based design scheme,which includes three parts:perception,decision and control.The rule-based design scheme are high security and good interpretation,but it's very expensive,highly complexity and needs high precision map support.The other is based on the end-to-end deep neural network design scheme,which is still in its infancy.Although its low cost,low engineering complexity and does not require high precision maps,but it needs lots of data,highly algorithm and not interpretable enough.This thesis mainly studies the key problem in the two design schemes of AGV tracking control method.By analyzing and summarizing the previous control methods,a finite frequency domain adaptive H? lateral tracking controller is designed.Also,the End-to-End deep learning control scheme is improved,besides,two controllers are simulated and compared.The details are as follows:Firstly,this paper analyzes and introduces different design schemes and modules of AGV,clarifies the functions of its main sensors,analyze and introduce the research and development of AGV and common control methods at home and abroad.The kinematics and dynamics modeling of the AGV is carried out,and the key parameters are identified.Secondly,it shows that when the actuator fails,the accurate detection of the front wheel declination can't be accurately detected by the existing sensors.Therefore,an adaptive mechanism together with finite frequency H? control strategy is introduced to estimate actuator fault and reject the effects of disturbances,and compared with a full frequency domain controller.Then,for the End-to-End deep learning control scheme,the basic knowledge of deep learning is introduced,and a CNN model is designed based on the NVIDIA model.After that,the Udacity simulator is used to obtain the data set required by the CNN model.After the training is completed,the simulator can be successfully run,and the End-to-End control scheme has a more intuitive feeling.Then MATLAB is used to simulate the finite frequency domain controller and the entire frequency domain controller.It can be seen that the finite frequency domain controller has better suppression effect on interference,its steady state error is smaller,and it is more reliable than End-to-End.However,considering the cost reasons and the current status of the research,the two control schemes are not opposite.How to use them together will be the focus of the next study.The conclusions and perspectives are presented in the end of the thesis.