Machine Learning Based Wifi Indoor Positioning Algorithm Research

With the advancement of technology and the improvement of people’s living standards, providing high quality services based on location becomes the general trend. Affected by complex interior structure, multipath propagation,pedestrian movement and many other factors, indoor positioning technology hasn’t been applied to reality in large scale, but rather still in the research stage.Indoor positioning based on Wi-Fi, to take full advantage of widely deployed wireless access in the building equipment and intelligent mobile terminals,became the focus of attention in the industry.Rapidly developing study of machine learning techniques infuses new vigor for Wi-Fi based indoor positioning. This paper seeks for methods to increase the positioning accuracy based on EWKNN algorithm and the Gaussian mixture model based fingerprint algorithm.This paper researched the Rayleigh mixture model based fingerprint positioning algorithm. The core idea is converting the received signal strength to received signal energy, so the distribution becomes even more concentrated.With limited samples, this method can better fit the model of probability distribution. As a consequence, the positioning accuracy is improved. In order to validate the method, a complete positioning system is designed and implemented.In order to apply the Rayleigh mixture model based fingerprint positioning algorithm in large target areas which are cross-building and across-floor,it is necessary to take into account the characteristic of sparse data which is collected in such area. To that end, this paper designed a set of pretreatment process and verified it on a large public dataset.Finally, this paper researched the multi-task based deep neural network positioning algorithm. The core idea of this algorithm is that, in the cross-building and cross floor environment, tap the floor and floor classification task correlation with specific coordinates regression task, so as to improve positioning accuracy. It’s proved that the multi-task based deep neural network positioning algorithm outperforms the Rayleigh mixture model based fingerprint positioning algorithm in terms of accuracy on the same large public dataset.