Research On Fingerprint Localization Technology Based On Fusion Of WIFI And Visual Images In Changeable Indoor Environment

With the development of the Internet of Things and artificial intelligence technologies,the importance of location-based services in the construction of smart cities is increasingly prominent,among which indoor positioning technology is the key to implementation.WiFi and visual signals have the advantages of rich information and low cost,and fingerprint positioning systems based on these two signals have become a research hotspot in the field of indoor localization.However,a series of emerging scenarios in smart cities often have the characteristics of frequent environmental changes,resulting in fluctuations in the positioning signal observations,which seriously affects the accuracy and reliability of positioning systems.Therefore,it is urgent to study how to use WiFi and visual images to achieve high-precision and high-robustness indoor localization.This paper addresses the issues of insufficient accuracy in single signal source positioning,high costs associated with acquiring data for multiple spatio-temporal positioning signals,and inconsistent distribution of fingerprint features due to signal fluctuations across different time and space conditions in changeable indoor environments.To leverage the complementarity of heterogeneous signals for achieving high-precision fusion location,an efficient fusion method of WiFi and vision based on low rank factor is proposed.To effectively expand multiple spatio-temporal real data sets,a cross space-time generation network based on location information consistency is presented.To improve the generalization ability of the positioning system in changeable indoor environments,a highly robust fingerprint localization method based on domain adversarial learning is proposed.The specific researches conducted in this paper are as follows:1.To address the problem that WiFi and visual fusion representations are difficult to fully exploit the complementarity of position information in two heterogeneous signals to improve the degree of position discrimination,this paper attributes the utilization of complementarity to the modeling of the specific effects of individual signal and the interactions between different signals,and a WiFi and Vision Fusion Method based on Low Rank Factor(LRF-WiVi)is proposed.Firstly,two feature extraction subnetworks are designed to extract the location information contained in the channel state information(CSI)and the multi-directional visual images,respectively,and achieve feature homogenization.Then,the fusion module based on low rank factor efficiently aggregates the specific actions and interactions of the two feature vectors while maintaining low computational complexity,resulting in a fusion representation with high location discrimination.2.To address the problem of high cost of field data collection and uneven distribution of different fingerprint data points,this paper proposes a Cross Space-time Generation Network based on Location Information Consistency(CSGN-LIC).CSGN-LIC learns to map real data samples under one spatiotemporal condition to another different spatiotemporal condition in the absence of paired training samples.In CSGN-LIC,the adversarial loss controls the transfer of sample styles between different spatiotemporal conditions,while the location information cycle consistency constrains the position information contained in the generated samples to ensure the effectiveness of data expansion.According to the characteristics of data samples,the specific structures of generation network and discriminant network are designed for WiFi channel state information and visual images respectively,and two kinds of cross-space generation network models suitable for different signals are obtained to achieve data expansion.3.To address the problem of reduced system positioning performance due to inconsistent fingerprint feature distribution in changeable indoor environments,this paper proposes a highly robust fingerprint localization method based on domain adversarial learning.The high robustness fingerprint localization in changeable environments is defined as a domain generalization problem,and a Multi Space-time Domain Adversarial Network(MSDAN)is constructed.The network consists of a fusion fingerprint feature extractor based on low-rank factors,a location estimator,and a domain discriminator.Through adversarial training on the multi spatiotemporal dataset constructed by CSGN-LIC,the extracted fusion fingerprint features are both domain-invariant and position-discriminatory,effectively improving the generalization ability of the positioning system in a changeable environmentAfter experimental testing and evaluation,LRF-WiVi proposed in this paper effectively improves the accuracy and stability of the fusion positioning system;CSGN-LIC enriches the sample quantity and diversity of the multi spatiotemporal dataset,thereby enhancing the positioning performance of the system,and addressing the problem of uneven distribution of reference point data under different spatiotemporal conditions to a certain extent.Compared with transfer learning and domain adaptation methods,MSDAN can more effectively improve the generalization ability of the positioning model in unknown spatiotemporal domains.The high robust positioning system based on the proposed algorithm has the mean positioning error of 0.69 m and standard deviation of 0.74 m in the changeable laboratory environment,showing excellent and stable performance.