Indoor Positioning Technology based on Deep Neural Networks

With the increasing demand for location based services, indoor positioning technology has become one of the most attractive research fields. Because the large indoor scenarios exist a large number of Wi-Fi signal, the researchers make it possible to use the existing infrastructure for locating in indoor environment and greatly reduce the cost in recent years. However, this positioning scheme also exists serious problem, the change of the indoor environment can affect the results of positioning, it will lead to the instability and low accuracy of positioning system, positioning method based on Wi-Fi has not been widely used at present. This paper proposes an indoor positioning solution based on the deep learning networks with the fusion of multi-source data, we design the deep learning networks model and introduce Restricted Boltzmann Machine to initialize networks and use grid search to optimize the parameters of networks. When data set is built, we consider the problem of whether the APs are moved or closed, this paper also considers the geomagnetic data, iBeacon in indoor environment, thus the dimensions of the input data are increased, and the stability of the indoor positioning system is improved, the method of cross validation fully trained network is utilized in this paper, at the same time, a data smoothing processing based on Kalman Filter (KF) is introduced. The result of the experiment shows that, average positioning error of scheme based on DNN and Kalman filter is 0.29m, the maximum position error is 1.59m, the position error is 96.33% within 1m, position accuracy is 100% within 2m, basically meets the positioning requirement.