| With the development of intelligent mobile devices,the application of location-based services is becoming more and more popular.The high-precision and robust indoor positioning technology is necessary for the further development of Internet of things and smart city.In outdoor environment,GNSS provides high-precision positioning services for many aspects,such as production and daily life.However,in the indoor environment,considering the building occlusion,complex obstacles and multi-path influence,positioning accuracy is restricted.So,it is urgent to carry out the research on high-precision and reliable indoor positioning technology to meet the needs of accurate and real-time indoor positioning services.The CSI in indoor environment is faced with three problems:the signal is greatly affected by noise,the resolution of amplitudes is insufficient as the area expend,and the phase distortion makes it difficult to extract features.Based on the analysis of the distribution of CSI amplitude feature,a self-iterative amplitude feature extraction algorithm is proposed;based on the analysis of three kinds of correlation among phases,a phase feature extraction network consists of multi-dimensional correlation is proposed;the final location is calculated by combining these two extracted features.The specific research work is as follows:1.The amplitude of CSI data is easy to be interfered by environment and hardware noise,and the resolution is insufficient with the expansion of the test area.This paper proposes a hyper-resolution adaptive feature extraction algorithm named clustering map(C-MAP).C-MAP consists of two parts:dynamic denoising and feature enhancement.In the dynamic de-noising part,in order to reduce the influence of empirical parameters,core features are obtained by iterative clustering.In order to ensure the adaptability,the clustering method is enhanced to effectively deal with extreme cases.In the feature enhancement part,fine-grained features are extracted by combining regularization polynomial fitting and curve gradient.Then,stable and high-resolution amplitude features are obtained by nonlinear kernel mapping.2.Considering the time correlation introduced by adjacent packets,the phase correlation between antennas and the subcarrier correlation,this paper innovatively proposes an extraction network of CSI phase named convolution graph convolution network(C-GCN).In C-GCN,the correlation in non-Euclidean space and the correlation in Euclidean space are considered simultaneously.The purpose of this network is to integrate the dynamic change into fingerprint database,and fully mine the correlation between antenna and subcarrier.In the graph convolution layer,C-GCN regards each sub-carrier of each antenna as a node in the graph,and constructs the edge connection by the correlation between antennas and sub-carriers.In the convolution layer,C-GCN extracts the correlation of time dimension with natural structure relationship through Euclidean space convolution.By combining the graph convolution layer and convolution layer,using the end-to-end training to calculate the new phase features including three kinds of correlation.3.For the bi-modal feature of CSI,a system named MAP-GCN is propose in this paper,which combine the stable hyper-resolution amplitude feature from C-MAP and the new phase feature including correlation from C-GCN.These bi-modal new features are stored in the fingerprint database to complete the final location calculation.In summary,the C-MAP algorithm proposed in this paper effectively improves the stability and feature resolution of CSI amplitude features;the-GCN network shows excellent performance of restraining error long tailing;the combined bi-modal positioning system MAP-GCN achieved an average error of 0.99m in indoor comprehensive environment and 1.14M in underground garage environment.MAP-GCN improves the accuracy of positioning in the range of 0-2m,and at the same time,it effectively suppresses the long tail error and improves the positioning stability. |
PDF Full Text Request