Unsupervised Localization by Learning Transition Model

Abstract

Nowadays, it becomes very convenient to collect synchronized WiFi received signal strength and inertial measurement (RSS+IMU) sequences by mobile devices, which enables the promising solution to conduct unsupervised indoor localization without the pain of radio-map calibration. To relax the needs of floor-map information or trajectory knowledge, this paper proposes to learn a transitional model (TM), which segments the massive unlabeled sequences to train a model that captures the expected relationship between {zt--1, zt } and ut--1, where zt--1, zt are two consecutive signal states at t and t -- 1, and ut--1 is the one step motion calculated from inertial data. We present both a transitional model in signal space (TMS) and a transitional model to predict motion from signal change (TMM) to represent the relationship in different ways. In particular, from the massive sequences, both the signal states and the one step motion are smoothed from the nearest neighbours, so that the transition model learns the expected relative signal state change triggered by the smoothed one step motion. Its distinctive features are that (1) no external floor-map or trajectory knowledge is needed; (2) it can be continuously on-line refined as unlabeled sequences are incrementally collected. KALMAN filter based on-line mobile user location tracking methods are given for both models. Experiments show that the transition model based localization method provides comparable accuracy with the manually fingerprint calibration methods.