Large-Scale Deep Learning Framework on FPGA for Fingerprint-Based Indoor Localization

Chang Liu,Chun Wang,Juan Luo

doi:10.1109/access.2020.2985162

Chang Liu, Chun Wang + Show 1 more

Open Access

https://doi.org/10.1109/access.2020.2985162

Copy DOI

Abstract

Localization Based Service (LBS) has become as one of the most important applications in modern daily life. Positioning technologies for outdoor environments are relatively mature because of the wide coverage of satellite navigating systems such as the Global Positioning System (GPS). In contrast, indoor localization remains a great challenge due to the fluctuated radio propagation environment. In addition to the high requirement of accuracy, an indoor localization system should also be low cost, low power consumption, and ubiquitous availability in mobile devices. To this end, fingerprint-based indoor localization schemes have served as an effective methodology to satisfy those requirements and have attracted more and more research attentions. In this paper, we present a scalable Deep Neural Network (DNN) architecture with Denoising Auto-encoder for Fingerprint-based Indoor Localization (called “SDNNLoc”) based-on FPGA implementation. First, a scalable stacked denoising auto-encoder is introduced to extract features from the fingerprint database for robustness and accuracy. Then, a generic parameterized DNN accelerator generating & optimization framework is proposed for FPGA implementation. In addition, we also demonstrate a WiFi-based fingerprinting indoor localization system for a crowdsensed university campus scenario. The experimental results show that the proposed DNN framework and its FPGA implementation are feasible for efficient and accurate indoor localization with good performance and high scalability.

Highlights

With the remarkable developments in Internet and mobile devices, Location Based Service (LBS) has permeated into numerous aspects of modern life
Inspired by the above research, we propose a scalable and accurate WiFi fingerprint location recognition method based on a Deep Neural Network (DNN), along with optimization methods on FPGA platform, to improve the accuracy and performance of a multistory WiFi location scheme, that uses: (1) a modified particle filter denoising algorithm based on sampling optimization, and (2) an automatic fingerprint update mechanism, which takes advantage of the continuity of motion, to solve the challenges brought by the environmental diversity, and (3) FPGA acceleartion for the proposed DNN algorithm (SDNNFIL) on the purpose of scalability and real-time operation
The Stacked Denoising Autoencoder (SDAE) can capture nonlinear and implicit correlations between fingerprint and the spatial location through signal changes at different reference points, which enables the trained network to capture the statistical dependencies among different inputs

Summary

INTRODUCTION

With the remarkable developments in Internet and mobile devices, Location Based Service (LBS) has permeated into numerous aspects of modern life. When the indoor LBS is expected to cover a large area like university campus or big shopping malls, scalability of the algorithm and its implementation will become one of the top priority Tackling on this problem, in this paper, we present a deep neural network with FPGA implementation for fingerprint-based indoor localization. THE OVERALL FRAMEWORK OF THE INDOOR LOCALIZATION Fig. presents the overall framework of the indoor localization which is based on the WiFi RSS fingerprints It consists of 3 stages: the offline database-building phase, the online location prediction & target tracking phase, as well as the online fingerprint updating and positioning phase. The database is updated on the basis of spatial correlation as a feedback mechanism to handle the environment dynamics

TRAINING FINGERPRINT USING DENOISING AUTO-ENCODER

EXPERIMENTAL EVALUATION

Findings

CONCLUSION