Autoencoder Matrix Completion Based Indoor Localization

The widespread of mobile devices facilitated the of many new applications that provide services based on user's location. Several techniques have been presented to enable such a service even in indoor environments where Global Positioning System (GPS) has low localization accuracy. These methods use some environment measurements. The most popular are using Received Signal Strength (RSS) for user location estimation. Due to the propagation conditions in indoor environment, the RSS methods suffer from missing data problem where the RSS can be below the sensitivity of some receivers. To overcome this problem, we propose in this paper an RSS matrix completion strategy based on an autoencoder algorithm as a preprocessing step. This latter exhibits a good performance in data denoising problems and can be applied for matrix completion purpose. A neural network is then used on the recovered RSS matrix to estimate a user's position. The performance of the proposed scheme is evaluated in a simulated environment and compared with traditional method of matrix completion based on the gradient descend algorithm and its variant. The results show the outperformances of our system of between 1 and 3 meters gain on localization error.