Research On Indoor Fusion Localization Technology Based On Channel State Information And Image

With the informatization and intellectualization of society,the role of information such as navigation and positioning has become increasingly prominent in daily life.Location-based service has been widely used in various fields.Aiming at the complex indoor environment,a variety of positioning methods have been proposed.Among these methods,Wi-Fi and visual signal have attracted much attention due to their advantages of rich positioning information and low hardware cost.In this paper,aiming at the limitations of single sensor for positioning,the indoor fusion location technology based on Wi-Fi Channel State Information(CSI)and visual image features is studied.CSI and image are two heterogeneous data sources for localization.In order to extract key features of the two localization sources and isomorphic the features,a CSI amplitude feature extraction network based on local correlation and a unified representation network based on multi-directional image features are proposed respectively.In order to realize the fusion of CSI features and image features,a fusion characterization method based on measure indicator for fingerprint construction is proposed.And the location matching and prediction is performed based on the method.The specific research works are as follows:1.Aiming at the problems of low matching degree between existing feature extraction algorithms and CSI amplitude and the insufficient feature discrimination,this paper proposes a CSI amplitude feature extraction network on basis of local correlation theory derivation:Local Connection Based Deep Neural Network(LC-DNN).In order to mine location-related information between adjacent sub-carrier CSI amplitudes,the local connection layer in LC-DNN uses multiple non-shared convolution kernels to extract local features from CSI amplitudes in different frequency ranges.In order to obtain the location-related global information,the full-connection layer in LC-DNN mainly extracts the target-related global features from the splicing data of various local features.2.Aiming at the high data dimension in image feature extraction and lack of disordered unified representation of multi-directional images,this paper proposes a unified representation network of multi-directional images:Shared Convolutional-Neural-Network Based "Add" Fusion Network(SC-AFN).In order to achieve uniform dimensionality reduction and characterization of images in different azimuths,the weighted shared CNN model is used to extract the uniform features of all images.After the feature extraction of all azimuth features,the "Add"strategy is adopted to realize the multi-feature fusion in network,which aims to reduce the amount of computation and realize the disordered fusion of all azimuth features.Finally,the full connection layer is used to extract and integrate the position correlated features of additive fusion data,which constructs the unified representation of the multi-directional image.3.Aming at the problems of insufficient fusion depth and lack of unified representation of heterogeneous features in the construction of fusion fingerprint database,this paper proposes a fusion representation and matching localization method based on measure indicator,and further improves the feature discrimination by introducing measure indicator to the model optimization process of fusion representation.Firstly,a perceptron-based fusion representation initial model is proposed to get the fusion representation domain from the above two localization features.Then the discrimination of fingerprint database is quantified as a measure indicator,which can be used for subsequent fusion guidance.Finally,the parameters of fusion representation model are optimized by maximizing the measurement indicator of the fusion fingerprint database,and the fusion representation model with high discrimination can be obtained.In summary,the proposed LC-DNN network can effectively improve the overall positioning accuracy and stability based on CSI amplitude,SC-AFN network reduces the maximum positioning error,decreases the probability of large error and improves the positioning accuracy.Based on the two feature extractions,the fusion fingerprint matching positioning achieves average positioning error of 0.62m and a standard deviation of 1.55m in indoor comprehensive office scenes.Compared with the LC-DNN and SC-AFN mentioned above,it improves the positioning accuracy by 16.2%and 50.4%,respectively.Compared with the existing Multidimensional Scaling-K Nearest Neighbor(MDS-KNN)fusion fingerprint matching,it improves the positioning accuracy by 11.4%.