Research On Indoor And Outdoor Navigation Scene Recognition Technology Based On Machine Learning

With the continuous development of civil aviation industry and the emergence of concepts such as smart civil aviation and smart airports,airport terminals are gradually developing towards functional,integrated,and intelligent directions.The footprint of terminal buildings is becoming larger and the internal spatial structure is becoming more complex.Passengers often spend a lot of time from arriving at the terminal to boarding the aircraft.Seamless indoor-outdoor positioning and navigation services can play a critical role in helping passengers plan their travel and improving the airport’s service level and passenger experience.The fundamental technology for this is indoor-outdoor navigation scene recognition,which is of great significance for selecting reliable positioning and navigation technology based on accurate recognition of the current navigation scene.This thesis proposes an indoor-outdoor navigation scene recognition technology based on smartphone sensors.By using the rich sensor sources embedded in smartphones,such as light sensors and magnetic sensors,the navigation scene features of different sensors are mined to construct a machine learning navigation scene recognition model.The model is trained and predicted using real data collected from various target navigation scenes.The main work is as follows:Firstly,the differences in light intensity,geomagnetic intensity and satellite information in different navigation scenes are analyzed,preliminarily proving the feasibility of the technology.In order to ensure the applicability of the technology,a large amount of rich sensor raw data was collected in different navigation scenes of the airport terminal and statistical feature data was extracted through data cleaning,feature extraction,feature selection and other operations to establish three data sets of light intensity,geomagnetic information and Bei Dou information used for training and prediction of the indoor-outdoor navigation scene recognition model.Secondly,the established data sets were divided into training set and test set,and five machine learning algorithms including naive Bayes,random forest,KNN,Adaboost,and Stacking were used to construct the recognition model.The established training data set was input into the recognition model.Through comparative analysis of experimental results,Adaboost had the highest recognition model performance,with performances of 85.1%,88.5%,and 92.4% on the three test data sets.To further improve the recognition accuracy,an improved Adaboost algorithm based on random forest was proposed,and the accuracy of the improved model was improved by 2.8%,2.7%,and 0.7%,respectively.Finally,the limitations of model discrimination performance based on single sensor data were analyzed,and a multi-sensor navigation scene recognition technology based on DS evidence theory was proposed to reduce the impact of environmental changes on single sensors.Through experimental analysis,the recognition results of the three sensor recognition models fused together in the test navigation scene can achieve an average accuracy of 98%,which is an average improvement of more than 7% compared to the machine learning recognition model based on a single sensor for indoor-outdoor transition navigation scene recognition.