PerLoc: Enabling Infrastructure-Free Indoor Localization with Perspective Projection

With the rapid development of mobile applications, there is an urgent need for highly efficient indoor localization service. Dedicated systems achieve good accuracy at the cost of deploying special hardware. Fingerprint-based methods avoid maintaining the expensive infrastructure but suffer from intensive labor for site-survey and poor robustness. In this paper, we present Per Loc, an infrastructure-free localization system which leverages rich vision features in indoor environment with high efficiency and accuracy. Per Loc makes use of binocular ranging technique to calculate the depth of a feature point and then figures out its geographical coordinates to build up reference point database, for which we design a filtering scheme to keep the database efficient in storage and search delay. During the localization stage, users simply take a photo of surroundings and feature points are extracted automatically as input for search scheme. Then a fast two-stage search scheme is proposed to find the nearest neighbors of query feature points in reference point database. Based on the perspective projection model, we inversely calculate users´ geographical location in real time. We implement the proposed localization system on commercial smartphones as well as laptops and conduct extensive experiments. Per Loc achieves 1.76m of average error in office environment, and 2.2m of average error in shopping mall.