Configurations And Analysis For Automatic Driving System Of Grand Theft Auto V

With the development of computer hardware,the enhancement of computing power,and the growing prosperity of the game market around the world,major game makers continue to develop games that are more realistic in both image rendering and physical simulation effects.On the other hand,it is benefited from the abundant image informa-tion that can be extracted from the game engine,as well as reliable label information.Some innovative technology institutions have gradually paid attention to the problem of using games to train machine learning algorithms.The Grand Theft Auto V(GTAV)published on the personal computer platform by Rock Star in 2015,has an abundant and authentic world design,which can provide a large and complex scene suitable for the training and testing of vehicle autopilot algorithm.Since GTAV is a commercial software,it is a closed source project.Although the internal material and dynamic system are proved to be externally controlled,there is also a lack of related documents.In order to transform the GTAV into a suitable automatic driving training field.The relevant research of this paper carried out on the basis of open source code,is mainly embodied in two aspects.One is to refactor and expand the content of open source code,add an additional virtual sensor-a completely configurable two-dimensional/three-dimensional laser radar;and two is to study the automatic driving system provided by GTAV for in-game vehicles.Due to the closed source nature of the GTAV,the auto-driving system provided by the GTAV is a black box system.But it allows two levels of control outside.The lower level is the directive control of a vehicle's throttle,brake and steering wheel.The higher level is the control of the driving behavior intention using four driving behavior parameters.The content of the work is divided into two aspects.One is to achieve the interactive data chain between the GTAV vehicle automatic driving system and the external programs and algorithms.On the higher level control part,the plug-in designed in this paper can make a GTAV vehicle driving automatically in accordance with preset driving behavior parameters,preset routes,and in total three ways to carry out,namely,one-way trip,cyclic one-way trip and cyclic round trip,so that the virtual sensors attached on the vehicles can sample the driving and environmental data in the GTAV.Two,by analyzing the four parameters necessary for automatic driving system configuration(called driving behavior characteristic parameters in this paper),the re-lated scheme is completed.Using these schemes to configure GTAV vehicle automatic driving system,this paper analyzed the relevance between these parameters and driving behaviors of the GTAV vehicle.In order to better reflect the distribution of probability density under different random degrees in the sequence of steering wheel angle(It is considered in this paper that the random degree is larger in the case of avoiding the sudden appearance of the obstacle ahead,while car following is smaller.),we propose a method in this paper for selecting the window width of kernel density estimation,which is based on the random degree.This method is based on a common window width selection method-Rule of Thumb,represent the degree of randomness by the correlation coefficients,and use the correlation time derived from the correlation coef-ficient to solve the uncertainty of the first order Gaussian Markov process,and finally use this uncertainty to replace the standard deviation of sample,which is required for Rule-of-Thumb window width selection method.The actual calculation process shows that the algorithm can obtain the sequence distribution under different random degrees by adjusting the correlation coefficient.