Application Of Artificial Intelligence In Navigation Positioning

Navigation positioning has great application value and prospects in the fields of driverlessness and public transportation.With the popularization of mobile intelligent computing platforms,civilian navigation software has gradually penetrated into the people's lives.Positioning accuracy is an important factor affecting the navigation software experience.However,the navigation system has limited positioning accuracy,and it is difficult to meet the daily precise navigation requirements by relying on the original positioning data.In order to solve the pain point of this application,this paper proposes the navigation and positioning adsorption problem firstly,and shifts the navigation position from the“point”to the the“line”,aiming at improving the accuracy of the starting position in the route planning.It solves the problem that the initial navigation position positioning accuracy is low in the navigation planning,which lays a foundation for the subsequent reasonable route planning.In view of the confidentiality of road network,this paper simplifies the research to abinary classification problem about whether a certain road is the real road position when the navigation is initiated.In the course of the experiment,referring to the idea of map-matching,combined with the GPS information when originating a navigation,assistant GPS points information before the navigation,the road information around the GPS point when originating a navigation,and the road network information,a total of 218 features are generated.One-way-ANOVA and two-way-ANOVA shows that at 95%confidence level,89.5%of the features are able to distinguish positive and negative samples separately,and 9.1%are able to distinguish positive and negative samples when other features are combined.In the experiment,the sample size is 21100,and the ratio of positive and negative samples is close to 3:2,which is approximated as equilibrium data.The SVM,RF and XGBoost algorithms are used to perform based on the five-fold cross-validation.The results show that,in view of the sample as a whole,the highest prediction accuracy of the SVM model based on the rbf function is 76.44%,and the training time is 9.74s.The highest accuracy of RF based on the Gini impure is 80.01%,and the model training takes 1.45s.The highest accuracy of the XGBoost model based on logistic regression is 82.61%,and the training is 8.48s.The longer training time of the XGBoost model has a greater relationship with the server performance.Due to limited resources,this article only runs programs under dual-core servers,so it does not highlight the advantages of XGBoost supporting parallelism.For the positive and negative samples,the accuracy,recall rate and F₁ value of SVM algorithm are extremely unstable;Results of RF algorithm are relatively stable,whose accuracy,recall and value are maintained,but the negative sample indicators are all lower than the positive samples;the XGBoost algorithm is consistent with RF,whose predictions for positive and negative samples are relatively stable.The indicators of negative samples are lower than the positive samples as SVM,but the accuracy,recall rate and F₁ value can be maintained at around 85%.In the experiment,the status of the negative is lower than the positive is related to the made-up samples.In addition,comparing the RF and XGBoost algorithm Top10 importance characteristics,the result shows that 60%of the features overlap.The high-precision distribution feature of the assistant GPS in RF is not an important feature expected.According to common sense,this feature should not appear as an important feature.The important features of the XGBoost algorithm are consistent with the expectation.It is a more accurate sorting of important features.The specific sorting results are:road length,phone direction angle,starting point flow,vertical dimension between the navigation point and the road,and the distance from the foot to the u pstream exit of the road,navigation point accuracy,connected road length,maximum of assistant GPS phone direction,average of assistant GPS phone direction,formway.In summary,the XGBoost has the best predictive power in this paper,and the accuracy of the binary classification is82.01%.In this paper,three different types of model are used for the binary classification.The model algorithm ranges from simple to complex,from weak learner to ensemble learning.The algorithms cover convex quadratic programming,decision tree and boosting tree.While gradually increasing the complexity of the model,reducing the variance and bias of the model,and achieving the gradual improvement of accuracy,it is an effective model selection scheme.However,the shortcoming is that the sample size is relatively insufficient,and it is impossible to realize sub-city and GPS number graded prediction to further improve the prediction accuracy.