A Research On Fingerprint Indoor Localization Based On WLAN

The popularity of Internet of Things have enabled the need of accurate indoor location tracking of users.Wi-Fi based indoor localization has attracted great interest due to its ubiquitous access in many indoor environments and the application of intelligent terminals.Locations of mobiles in a Wi-Fi based indoor localization system can be determined by received Wi-Fi signal strength.Machine learning has been widely used to achieve good accuracy in meters.Signal strength variation caused by multiple factors,which makes the localization model inaccurate.In summary,time and devices can be concluded as the chief factors of the characteristics of signal variation.The accuracy is deteriorated by the complex indoor propagation environments,which result in variable received signal strength(RSS).In addition,the difference in signal strength between different devices is the key problem of fingerprint positioning.When the terminal used online is inconsistent with the terminal used off-line,the positioning accuracy will be damaged due to the heterogeneity of the devices.The manual collection of fingerprints requires a lot of manpower and material resources,which hinders the popularization of fingerprint location.The fingerprints stored in the database are the values of samples gathered in the off-line phase,whereas the values of RSS fluctuate because of the multi-path and shadow fading effect,which are influenced by the temperature of the indoor environments,the human movement and the direction of the terminal.Therefore,dealing with the RSS variation is the main challenge to improve the accuracy of the localization system.According to the instability of RSS signal,we propose a positioning model based on deep auto-encoder and integrated regression,which enhances the stability of high-precision indoor positioning.We firstly preprocess the noisy RSS.So that,the important features of fingerprints at each reference point can be effectively extracted while reducing noise.In the off-line training phase,we train the deep auto-encoder to denoise the measured data and then build the RSS fingerprints according to the trained weights.Secondly,it is time-consuming and laborious to construct fingerprints by manual collection.In this model,regression is adopted instead of classification to effectively ensure high precision and collect database with low density reference points.In the online localization phase,we adopt three machine learning algorithms,which are random forest regression,multi-layer perceptron classification and multi-layer perceptron regression,to estimate the location.Then the mapping relationship between fingerprints and coordinates can be established.The real-time RSS can be matched to the final estimated location by the learned system.Different devices use different types of wireless cards and antennas,which have different sensing capabilities.Hence,the distribution of RSS can be different.New RSS fingerprints should be collected again when the signal distribution changes in order to keep the localization accuracy,which is costly or infeasible in large and complex buildings.Moreover,mobile phone update iteration speed is so fast,the fingerprints database cannot contain all kinds of terminals on the market.In the online stage,no tags of mobile models can be labelled and it is impossible to know the model of the terminal used in the collection of the fingerprint database.A transfer learning-based framework that combines domain adaptation with the elimination of diverse diversities is proposed to enhance the scalability of fingerprint-based indoor localization.The second-order statistics of source and target distributions of fingerprints collected online and off-line can be aligned to minimize domain shift,without requiring any target labels.The alignment performs before the classifier training process.As the target domain is fingerprint collected off-line and the source domain is fingerprint collected online,whitened source features can be aligned with target features to release the damage caused by device diversities.All the experiments in this thesis were carried out in real indoor scenes.Experimental results indicate the proposed systems are reliable with very high positional accuracy,which had certain promoting effect on the progress of indoor positioning.