Vehicle Prediction Based On Gaussian Process Regression Under GPS Failure

Nowadays,vehicle localization has become one of the most fundamental challenges of intelligent transport systems(ITS),which aims at increasing the driving safety as well as enhancing the road efficiency.A combined data fusion method for vehicle localization consisting of GPS and Inertial Navigation System(INS)is one of the most promising technologies that is capable of enhancing positioning accuracy under challenging environment.In the integrated GPS/INS system,GPS-derived positions can be used to calibrate INS.On the other hand,INS bridges GPS outages and reduces the search domain required for detecting and correcting GPS cycle slips.To achieve reliable integration between INS and GPS,some algorithms are employed to provide an optimal estimation of vehicle navigation state.However,in the complex modern urban traffic environment,there are many problems such as GPS signal interference and sensor data fluctuation,which leads to the decrease of modeling difficulty prediction,which affects the realization of most application services in vehicle self-organizing network.Therefore,the realization of real-time,continuous and reliable vehicle location information in extreme urban traffic environment has become a hot issue in ITS field research.This paper mainly studies the vehicle location prediction problem in modern urban traffic environment.Based on GPS/INS integrated navigation system,multi-source data fusion is adopted to obtain reliable and reliable vehicle location information.The main works of this thesis are as follows:According to the characteristics of intelligent transportation system,the significance of vehicle positioning is expounded.Based on GPS and INS navigation system,the basic idea of vehicle position prediction and the factors influencing the prediction accuracy are described.By analyzing the traditional vehicle position prediction algorithm,it is found that they are limited,and are not suitable for the increasingly complex urban traffic environment.A Particle Swarm Optimization based algorithm for hyper-parameters optimization of Gaussian Process Regression is proposed(GPR-PSO).The algorithm is based on the vehicle history trajectory data training modeling,combined with INS data to achieve reliable vehicle position prediction.The experimental results show that the algorithm improves the accuracy of BPNN,SVR and PLSR by 22.8%-65.5%.A regression model of Gaussian process based on Adaboost is proposed.In the process of PSO optimization of GPR superparameters,it is still possible to fall into the local optimal solution,and the GPR prediction effect depends directly on the selection of kernel function.Therefore,Adaboost integrates GPR weak classifier to obtain strong predictor,improves the learning ability of GPR and Generalization ability.The experimental results show that the GPR algorithm based on Adaboost is 13.33%-58.06%higher than GPR-PSO algorithm.This paper is based on the Java programming language and Matlab platform to realize the data acquisition and the realization of the algorithm,which provides a simple and effective method for the realization and evaluation of the vehicle position prediction algorithm.