Deep Generative Model-based RSSI Synthesis for Indoor Localization

Abstract

Indoor localization via deep learning (DL) is attracting researchers' attention. DL is mainly used for fingerprinting-based indoor localization as it generally employs a vast offline database to ensure its reliability. However, the long effort and high cost of constructing this database are the disadvantages of this technique. This paper implements variational autoencoders (VAE), one of the popular deep generative models, to alleviate the drawbacks of offline database issues. Our proposal works using the received signal strength indicator (RSSI); unfortunately, it is known for its fluctuation and instability. Thus, instead of using RSSI directly as a localization parameter, we learn its distribution via VAE to generate the synthetic RSSI values. We utilized the RSSI from an actual measurement campaign. The VAE implementation results show that we can obtain the RSSI synthesis by exploring the latent distribution learned from the input distribution. Thus, the offline database density grids can be enhanced. We validated the results by varying epochs to map the learned latent distribution. However, we still have relatively low accuracy in the synthetic RSSI values, especially when applying a small number of epochs, i.e., 10 and 100. When we applied epoch number 1000, the error was relatively low (-3dBm average error) in the sampled position. Our preliminary assumption is that the dataset is small for VAE learning, and probably the 3-by-3 RSSI-to-image size assumption could still be inadequate.