| With the rapid development and widespread popularity of wireless communication,the services based on mobile terminals'location information play an increasingly im-portant role in people's life.Since the accuracy of satellite positioning system is limited by the shadow effect in the environment and the non-line-of-sight propagation condition,it is no longer applicable in indoor scenes.So the problem of finding high precision posi-tioning technology for complex channel environments needs to be solved urgently.Wi-Fi wireless systems have been widely deployed.On the other hand,positioning accuracy of geometric positioning technology needs accurate geometric parameter measurement,which not only has great demand for communication bandwidth,but also has great re-quirements on the performance of hardware modules of mobile terminals.The conditions are difficult to meet.Therefore,in order to proposes solutions and suggestions for high-precision positioning technology in complex channel environments,this paper focuses on researching the fingerprint-based Wi-Fi localization system.The paper uses the off-the-shelf Intel 5300 network adapter and modified driver to measure IEEE 802.11n channel state information.By analyzing various features to illus-trate that the feasibility of the experimental platform and the feasibility of channel state information as a fingerprint data which can be used for localization.In order to increase the number of channel state information features as much as possible which is helpful to improve the positioning accuracy,a method of using antenna one by one to measure CSI is proposed to simulate a virtual transmit antenna array.Starting from the real application scenario,this paper designs two channel state information collection schemes to study the influence of dynamic changes of environment on positioning accuracy.In order to meet the real-time requirements of the positioning system,principal component analysis is proposed for feature dimension compression of CSI data.And then we establish an of-fline fingerprint database.In this paper,two common fingerprint localization algorithms,K-Nearest Neighbor classification and neural network regression,are implemented.By analyzing the position error,metric learning is proposed to improve the distance between channel state information samples to reduce the error.In addition,multi-sample statisti-cal positioning and local outlier factor algorithms are proposed to reduce the variance of coordinate estimation error.By realizing the localization algorithm proposed in this paper and analyzing the po-sitioning results,we can conclude that the positioning accuracy of the neural network is higher than that of the K-Nearest Neighbor algorithm.In the two application scenarios studied in this paper,the average positioning error of the neural network is 1.60 meters and 8.54×10-2meters of KNN.In the stage of establishing the database stage,selecting more training points and increasing data sample diversity are effective ways to improve positioning accuracy.The distance metric learned by the large margin nearest neigh-bors algorithm helps to increase the distance between data samples belonging to different training points which helps reducing the average positioning error,which makes the po-sitioning accuracy is close to that of convolutional neural network based on large-scale Multi-Input Multi-Output.Multi-sample statistical positioning and local outlier factor algorithm can generally reduce the interquartile range of coordinate estimation under the condition that the average positioning error is not increased. |
PDF Full Text Request