A Research On Model Parameters Transfer Methods In Indoor Positioning

With the development and popularization of internet of things technology and portable terminals,location-based services gradually play an important role in people's lives.Global positioning system(GPS),which can provide high-precision location services around the clock,is the most mature positioning technology at present.However,satellite signals are easily interfered in complex indoor environment and can not provide reliable location information.Therefore,a large number of researchers have put forward multiple indoor positioning technologies based on different principles by using various signal sources such as bluetooth,infrared,WiFi,ultra-wideband and so on,among which the positioning technology based on WiFi signals has become the most promising technology in the field of indoor localization with its advantages of low cost and easy popularization.The main problem in indoor positioning scene is that RSS distribution is easily changed by environmental changes and heterogeneous devices,which leads to the original positioning model being unable to locate accurately in the current environment.Traditional positioning algorithms assume that the data obey the same distribution,which can't solve this practical problem.However,the commonly used domain adaptation techniques pay too much attention to the commonness between domains and can only reduce the domain differences insufficiently from the mean or variance of data distribution.Aiming at the defects of the above methods,and considering that the original positioning model contains abundant knowledge,this thesis proposes three parameter-based methods to overcome the influence of RSS distribution discrepancy.In this thesis,a deep ensemble indoor location algorithm based on parameter integration is proposed.In the algorithm,unlabeled target domain data is used to assist the source domain data to train a new location model,and double constraints of mean and covariance are imposed on feature representations in different domains.Compared with traditional deep domain adaptation methods,it can minimize domain differences efficiently.Considering the jitter of the output of neural network in the training process,the model parameters are integrated by exponential moving average mechanism,so that the model used for online positioning has a stable predicted output.Experiments show that,without updating the fingerprint database,the algorithm can adapt well to the influence of environmental changes and equipment differences,and has higher positioning accuracy than other comparison algorithms.In this thesis,a deep transfer method based on parameter prediction and a dualnetwork architecture method based on parameter constraints are proposed,which overcome the disadvantages of common domain adaptation technologies that use the same network to extract different domain features,that is,they focus too much on the commonness of the two domains and ignore the unique characteristics of their respective domains.The method based on parameter prediction learns a transformation matrix according to the parameters of the source domain network to predict the parameters of the target domain network,so that the target domain network can adapt to the new environment while retaining the processing capacity of the source domain network.The method based on parameter constraint uses two networks to process the data in the source domain and the target domain respectively,linearly compensates the domain drift from the perspective of model parameters,and can adjust the parameters of the two networks more flexibly,thus fully mining the data features in different domains.Experiments show that these two algorithms can not only overcome the influence of environmental changes and heterogeneous devices,but also achieve robust positioning in the case of fewer target domain samples.