Fingerprint Augment Based on Super-Resolution for WiFi Fingerprint Based Indoor Localization
Indoor localization utilizing WiFi fingerprints has become a central focus of research due to its cost-effective equipment deployment and high accuracy. Even though various mobile apps offer smarter services based on real-time user locations based on GPS, they are only effective outdoors, as indoor positioning faces challenges due to signal scattering and multi-path propagation.
Therefore, achieving low-cost, high-accuracy indoor localization is an important research domain. Among various indoor localization methods like WiFi, Bluetooth, and ZigBee, WiFi emerges as a research hotspot due to its high positioning accuracy and low material cost.
Fingerprint-based positioning relies on collecting received signal information from Access Points (AP) by User Equipment (UE) at Reference Points (RPs), creating a detailed fingerprint database. The process has two crucial phases: offline and online. During the offline phase, the system captures nuances by recording RPs' coordinates and signal information from each AP in the WiFi's received signal strength (RSS). This forms the foundation of the fingerprint database. In contrast, the online phase facilitates real-time positioning by comparing coordinate-lacking fingerprint data with the established offline database. As a result, fingerprint-based WiFi technology excels in indoor positioning due to its exceptional accuracy.
This paper introduces a super-resolution-based Deep Neural Network (DNN) to approximate the mapping function between sparse and dense fingerprint databases. The Fingerprint Augment Based on Super-Resolution (FASR) framework integrates super-resolution with WiFi fingerprint augmentation, involving key components such as WiFi fingerprint database to low-resolution image conversion, super-resolution for low-resolution images, and high-resolution images to WiFi fingerprint database conversion.
The performance of the FASR framework is influenced by the geographic spacing of RPs and the upsampling factor of DNN, which renders crucial parameters for effective and practical implementation of the setup. Results show improved fingerprint augment accuracy and positioning accuracy. It validates the effectiveness of FASR in both numeric simulations and practical testing.
The study also explores Gaussian process regression (GPR) models, particularly Matern & Rational quadratic with isotropic distance measure compound kernel (MR-GPR), demonstrating effective fingerprint augmentation.
Recent progress in deep neural network research has significantly improved fingerprint augmentation using advanced techniques like SRCNN (Super-Resolution Convolutional Neural Network), EDSR (Enhanced Deep Super-Resolution Network), and SRResNet (Super-Resolution Deep Residual Network). These networks are designed to enhance image resolution; in particular, SRResNet generates high-resolution images with smooth textures. The FASR framework built on these networks consists of three modules and offers comprehensive evaluation in both simulated and real-world positioning scenarios, showcasing consistent performance across various room sizes.
The FASR framework transforms fingerprint augmentation into a super-resolution task to enhance WiFi fingerprint-based indoor localization precision. Dedicated methods, Fingerprint-To-Image conversion, and Image-To-Fingerprint conversion are introduced for FASR implementation.
The geographic spacing of RPs and the upsampling factor of DNN considerably affect FASR's fingerprint augmentation and localization performance. While FASR performed admirably, its effectiveness hinges on factors like room sizes, RP distributions, grid and channel characteristics. Future work aims to refine the FASR method to address these challenges.