Research On Key Technologies For Scalable And Low-Overhead Indoor Localization

With the development of mobile applications,the requirement of location-based services(LBS)is increasing.Besides,the popularization and evolution of mobile terminals further drives the LBS application requirement,such as intelligent mobility and smart home,which have penetrated into every aspects of people's daily life.At present,the WiFi based method is one of the most promising solutions to indoor localization,but it faces several pratical problems.First,the site survey process of fingerprint based method during the offline training phase is time-consuming and labor-intensive,which hinders the deployment and promotion of the location system.Second,due to the various environmental interferences,the WiFi received signal strength indicator(RSSI)is fluctuant,which leads the anti-jamming of location system is poor and further affects the location accuracy.Third,the localization process during the online phase requires global search in fingerprint database which may result in high and redundant computation overhead.To address the above issues,this thesis is dedicated to propose new indoor localization system architecture and algorithms based on key technologies including transfer learning,stacked denosing autoencoder and annulus local search.The main contributions are summarized as follows:(1)Regrading to the problem that the site survey process is time-consuming and labor-intensive,the transfer learning is employed to reduce the offline training overhead.The fingerprint-based method requires a site survey process to collect WiFi RSSI data at each reference point and construct the offline database,which brings burdensome overhead.We propose a transfer learning-based framework to enhance the scalability of indoor localization system with the purpose of reducing offline overhead and maintaining the localization accuracy.The framework consists of two parts: metric learning and metric transfer.The metric learning process learns a distance metric from source domains with sufficient label data and builds a metric pool.The metric transfer process chooses a proper distance metric for new location environment and reshapes the logical distribution of target domain,which makes the distance between online point and reference point that belongs to the same cluster in physical space closer and vice versa.Comprehensive experiments in real environment demonstrate the effectiveness of the proposed framework.(2)Regarding to the problem that the WiFi RSSI is unstable due to various environmental interferences,the denosing autoencoder method is employed to build a robust fingerprint database and improve the localization accuracy.We propose a dual-band stacked denosing autoencoder based indoor localization model.Specifically,in the offline phase,we collect 2.4GHz and 5GHz frequency of Wi-Fi RSSI data to extend the feature dimension.Then we adopt the stacked denosing autoencoder neural network to train RSSI data to build the offline fingerprint database.In the online phase,we propose a data generation scheme to offset the insufficient offline data and enhence the input.Finally,we employ local weighted liner regression method to locate the target.The real-world experiments show the robustness of the proposed method compared with the-state-of-art approaches.(3)Regarding to the problem that conventional fingerprint-based methods require the global search of the offline database during the online phase,which may bring high computation overhead and reduce localization accuracy.We propose a scalable localization algorithm,which consists of two parts: annulus-based localization(ABL)and local search-based localization(LSL).First,we build a distance-RSSI model with multinomial function fitting method to measure the distance between router and mobile terminal.Then,based on WiFi RSSI signal attenuation characteristic,we design a subarea division scheme to choose proper distance-RSSI function.Besides,we propose an annulus construction scheme to refine the retrieval space.Further,we propose a local search scheme based on WiFi RSSI fingerprint mapping.Finally,with the modelling of WiFi RSSI signal distribution,the KNN is adopted to compute the coordinate of the target.The experimental results conclusively demonstrate that proposed algorithms can improve the localization accuracy and reduce the online computional overhead.