Research On The Graph-based Un-supervised Learning In Indoor WiFi Fingerprint Localization System

Location-based Service(LBS)attracts more and more attention with the rapid development of smartphones and mobile Internet applications recently.Real time localization has become the fundamental technique of multiple high level applications such as traffic,commercial services,logistic and personality services.In outdoor scenarios,a good performance has been provided by the Global Navigation Satellite Systems(GNSSs),such as Global Positioning System(GPS).In indoor environments,however,GPS cannot reach high accuracy due to signal fading and multipath effect.Therefore,indoor positioning has gained increasing interest and become a hot topic in the navigation field these years.WiFi indoor localization is one of the most widely used indoor localization scheme,because of the low cost,easy deployment,etc.WiFi indoor localization systems consist of two phases: offline phase for fingerprint training,and online phase for locating.Conventional WiFi indoor localization systems fulfill the offline phase by dividing the area of interest into normalized grids,collecting Received Signal Strength(RSS)vector at each reference point in a static way and acquiring location information of reference points by participant manual input.We define such method as supervised learning method.In a natural situation,the static sampling process brings a problem: heavy workload upon users for fingerprint training in the offline phase.To overcome this problem,we develop a new graph-based fingerprint training method using un-supervised learning.In the proposed method,priori information of the indoor graph is firstly built.Subsequently,sampling path from the Starting Point(SP)to the Turning Point(SP)is computed with the method base on SP matching and edge matching.Finally,the samples on the path are obtained by linear interpolating base on time stamp.Fingerprint database can be derived from these steps.Then the database can be improved for a higher positioning accuracy and a wider scope of application by RSS modeling and samples clustering.Since the procedure of generating fingerprint database can be completed simultaneously with daily activities,it is unessential to spend specific labor for fingerprint training,which significantly decreases the workload of fingerprint database building,and guarantee positioning accuracy at the same time.