# Hapi: A Robust Pseudo-3D Calibration-Free WiFi-based Indoor Localization System

## TL;DR

Hapi is a calibration-free WiFi-based indoor localization system that accurately estimates a user's floor and 2D position using deep learning and RSS-Rank Gaussian methods, with high accuracy demonstrated in real-world tests.

## Contribution

Hapi introduces a novel pseudo-3D indoor localization approach that leverages signal attenuation and deep learning without requiring calibration.

## Key findings

- Floor identification accuracy up to 95.2%
- Median 2D localization error of 3.5 meters
- Significant improvement over existing calibration-free systems

## Abstract

In this paper, we present Hapi, a novel system that uses off-the-shelf standard WiFi to provide pseudo-3D indoor localization. It estimates the user's floor and her 2D location on that floor. Hapi is calibration-free, only requiring the building's floorplans and its WiFi APs' installation location for deployment. Our analysis shows that while a user can hear APs from nearby floors as well as her floor, she will typically only receive signals from spatially closer APs in distant floors, as compared to APs in her floor. This is due to signal attenuation by floors/ceilings along with the 3D distance between the APs and the user. Hapi leverages this observation to achieve accurate and robust location estimates. A deep-learning based method is proposed to identify the user's floor. Then, the identified floor along with the user's visible APs from all floors are used to estimate her 2D location through a novel RSS-Rank Gaussian-based method. Additionally, we present a regression based method to predict Hapi's location estimates' quality and employ it within a Kalman Filter to further refine the accuracy. Our evaluation results, from deployment on various android devices over 6 months with 13 subjects in 5 different up to 9 floors multistory buildings, show that Hapi can identify the user's exact floor up to 95.2% of the time and her 2D location with a median accuracy of 3.5m, achieving 52.1% and 76.0% improvement over related calibration-free state-of-the-art systems respectively.