A Low-overhead Indoor Positioning System Based On TrAdaBoost Transfer Learning

With the rapid development of Internet technology,location-based services have been integrated into our lives and become an indispensable part of our daily lives.Due to the rapid update of mobile devices,many mobile applications have also embedded positioning functions.In the outdoor environment,the Bei Dou Navigation Satellite System,the Global Positioning System(GPS),the Galileo system and other positioning technologies based on satellite signals and radar have been very mature,and due to the characteristics of strong anti-interference ability,high precision,and good real-time performance,these positioning systems can already meet the daily needs of the people.However,in the indoor environment where people often exercise,GPS and other outdoor positioning systems cannot locate because the signal is blocked,so they are not suitable for indoor positioning research.As one of the indoor positioning solutions,the indoor positioning system based on Channel State Information(CSI)uses existing WIFI(Wireless Fidelity,WIFI)resources for indoor positioning.It has the characteristics of simple operation,no need to carry equipment,and high accuracy,which has attracted wide attention and study from researchers.The practical CSI-based indoor positioning system faces two main challenges: one is the severe multipath and shadow fading influences,and another is the vulnerability to dynamic environment.The short-term interference(e.g.door opening,closing and movement of tables,chairs and other furniture)and long-term interference(e.g.humidity,temperature and light changes)will cause signal unavailabe.Therefore,the real-time CSI data will be quite different from the values in the fingerprint library.If the fingerprint library is not updated in time,the positioning accuracy will be reduced.To adapt to the changes in the environment,a simple solution for indoor location is to re-collect data and supplement the fingerprint library.However,this method is very impractical because this calibration process is timeconsuming and laborious.In view of the high-consuming and low accuracy of indoor positioning in a dynamic environment,this dissertation proposes a channel state information passive positioning method based on TrAdaBoost transfer learning(Transfer Learning).The main work is as follows:(1)After analyzing the signal propagation model and the advantages in the wireless environment,through comparison,it is proved that CSI has time stability,frequency diversity and other characteristics compared with RSSI.Therefore,CSI is used as the fingerprint feature of indoor positioning,and the effective phase information of CSI is extracted through phase correction,which provides the basis for the following experiments.(2)The CSI data is preprocessed by Principal Components Analysis(PCA),and the high-dimensional CSI data is mapped to the low-dimensional space.And use the reduceddimensional data as fingerprint features,and respectively in the source domain and Generate a data set in the target domain.After analyzing the proportion of discrete points in the data,a correction factor is used to reduce the weight of the source domain to optimize the iterative process.(3)Use One-hot coding and One-vs-Rest algorithm to realize the multi-classification ability of TrAdaBoost algorithm,and improve the learning effect through multiple iterations,and finally transform the final output into a two-dimensional space coordinate through regression of confidence probability.We conducted experiments on an empty platform,two teaching laboratories,and an office in a total of four scenarios to evaluate the performance of the method.Compared with the method that does not use transfer such as Boost method,the accuracy of our proposed method can be increased by 35% in a dynamic environment,and Site Survey Overhead can be reduced by 40%.The comparison with the currently popular indoor positioning methods verifies the effectiveness of our proposed method.We compared with the other three transfer learning methods and found that the positioning accuracy of TrAdaBoost is better than other methods.