Research On NLOS Identification Of Acoustic Localization Under Indoor Environments

In recent years,acoustic indoor localization technology has become the most potential solution with the advantages of good compatibility with intelligent mobile terminals and low cost.However,the indoor environment is diverse and complex,and the phenomenon of non-line-of-sight is common,which makes it difficult to achieve high-accuracy localization.The localization accuracy and stability can be improved through the identification and processing of NLOS.Therefore,in this thesis,semi-supervised and unsupervised online learning methods based on acoustic channel characteristics are used to break through the technical bottleneck of NLOS identification and provide a new method for this field.The main contents of this thesis are as follows:(1)Based on the estimation of relative gain and time delay,the acoustic channel difference between LOS and NLOS paths is characterized,and the 11 features are extracted for NLOS identification.(2)A kernel principal component analysis method of relief feature selection is proposed,which is used to select sensitive features and extract principal elements,eliminate redundant and irrelevant features,and data dimensionality is reduction.Based on the support vector machine classifier,the results show that it can improve the identification efficiency and the accuracy reaches 95%.(3)A semi-supervised learning method is proposed to solve the problem that the cost of sample label is high when the supervised learning classifier is used to the acoustic channel NLOS identification.Its effectiveness is verified by simulation,and the optimal classifier and sensitive feature set are given.The results show that the method can reduce the size of labeled sample to 700 and still achieve a identification accuracy higher than 90%.(4)An unsupervised online learning method based on sample weighting is proposed to solve the problem that supervised and semi-supervised learning classifiers have poor stability in processing dynamic data.The effectiveness of the online learning method is verified by the simulation,and the optimal classifier is given.The results show that the method can update the classification model online to deal with the dynamic data,and the performance of classifi-er is stable and the identification accuracy is higher than 85%.