Data Collection And Reconstruction In Wireless Fingerprinting Localization System

With the advancement of technology,location-based services play a very important role in our life.People's mainly get their location by GPS,global positioning system,but this system could not provide high quality services in different scenarios.Some devices cannot get service since they are not equipped with GPS modules.In addition,GPS cannot guarantee on accuracy in urban areas and indoor environments.As a complementary tool,Wireless fingerprinting technology could provide stable service in these scenarios.It has gained much recognition due to its high accuracy and low price for deployment.It does not rely on any principles describing the signal between transmitter and receiver,but the fingerprint database which gives a mapping from signal strength to location information.Currently,a challenge in wireless fingerprinting is how to acquire a large amount of fingerprint data to construct the database.The mobile crowdsensing framework has provided a new method for fingerprints collection.However,the reliability of data cannot be guaranteed since the data collection process is probably not completed by professionals with special devices.A critical problem we have to pay attention to is how to accurately evaluate the quality of the reported fingerprint data.This article proposes a data pricing and prediction mechanism based on the online learning framework.By analyzing the nature of localization error occurred during the fingerprint database construction process,the quality of the reported fingerprint data can be accurately evaluated with error function.Based on the theoretical framework of online learning,we transform the error function into a corresponding loss function,and design an online pricing mechanism based on data quality.In addition,we discuss the effect of pricing distribution function on the proposed mechanism in practical scenarios.The experimental result has shown the effectiveness of our mechanism,which can extract enough information from fingerprint sequence even though the budget is insufficient.Our method could eliminate the uncertainty caused by the fluctuation in the fingerprint signal collection.Considering the cost and topographic factor,the data collection with mobile crowdsensing is still unable to cover all the locations in target region.We need to predict the fingerprints of the unknown region by sampled fingerprints,and reconstruct the complete fingerprint database.In response to this challenge,we present a radio map reconstruction scheme based on Gaussian Process Regression model.Gaussian process is a non-parametric model,which uses kernel function to search a huge solution space.Through analyzing the applicability of different kernels in Gaussian process to different signal distributions,we propose a method to construct the optimal kernel function according to basic kernel functions,so that the predicting results of Gaussian process could approximate the true signal distribution in target region.Based on fingerprints collected in actual cellular network,we designed the experimental scheme.By comparing the prediction errors of Gaussian processes with different kernel functions,we show the superiority of our proposed optimal kernel function in the Gaussian process model.The proposed fingerprints reconstruction scheme could avoid the need of manually collecting a large number of fingerprint samples while keeping the accuracy of the database,thus ensuring the performance of the fingerprint localization system.