The 13th International Conference on Indoor Positioning and Indoor Navigation (IPIN) will be held on September 25th-28th in Nuremberg, Germany.

IPIN is the premier annual scholarly venue in indoor positioning technologies, one of core foundations of ubiquitous computing. The following presentation will be delivered.

This paper is the results of joint research between Yahoo Japan Corporation and Kawahara Research Group (The University of Tokyo).

In our group, this research is mainly conducted by our alumni, Mr. Takuhiro Shimokawa, based on his master thesis.

Data-driven simulation of wireless communication signal strength in indoor environments

In the field of ubiquitous computing, modeling the spatial distribution of received signal strength indicator (RSSI) of radio waves is an important topic, because it can be used to optimize the placement of beacons (access points) and to maximize coverage in indoor positioning. Physical modeling of radio propagation (ray tracing) techniques have often been used to simulate RSSI data for beacons placed at arbitrary locations. However, Ray-Tracing models can only ensure the accuracy of simulations by reflecting physical characteristics such as reflection and transmission attenuation rates of walls and floors within a building. In addition, many studies have been conducted to collect RSSI data by scattering beacons in advance, and to model the spatial distribution of RSSI using machine learning such as Gaussian processes. However, when the beacon position is changed, the data must be re-collected, making it difficult to use the system for the original purpose of simulation: “temporarily placing beacons and generating RSSI data so as to examine optimal placement.”

In this study, we propose a data-driven RSSI simulation method as a completely different approach to physics-base simulations. In our approach, the RSSI spatial distribution is first collected from randomly scattered beacons, and then, through the generative neural approach described below, the spatial distribution of RSSI emitted from beacons placed at arbitrary locations can be generated with high accuracy. This research utilizes a Generative Adversarial Network (GAN) and U-net to learn the relationship between beacon placement and the spatial distribution of RSSI.

To evaluate the proposed method, we actually collected RSSI data in an office environment (approx. 750 m^2), and evaluated its accuracy using actual RSSI values. The results showed that the average absolute error of RSSI was improved by 8% compared to physics-based simulations. We also constructed an indoor positioning model obtained by machine learning using only simulated data without any real RSSI data (so to speak, machine learning with zero-shot training data), and showed that it improves the average positioning error by 19% compared to the conventional simulation approach.
—-
Presentation information (Program)

September 25th (Mon.) 15:30 – 17:00 A.4 Machine Learning for Localization and Navigation

Data-driven simulation of wireless communication signal strength in indoor environments

Takuhiro Shimokawa*, Kota Tsubouchi†, Yoshihiro Kawahara‡, Hiroaki Murakami‡, Masamichi Shimosaka*

*: Tokyo Institute of Technology
†: Yahoo Japan Corporation
‡: The University of Tokyo