Towards Robust WiFi Fingerprint-Based Vehicle Tracking in Dynamic Indoor Parking Environments: An Online Learning Framework

Abstract

The variation of wireless signal in dynamic indoor parking environments may seriously compromise the performance of fingerprint-based localization methods. In this regard, this paper investigates the problem of robust WiFi fingerprint-based vehicle tracking in dynamic indoor parking environments, aiming at designing an online learning framework to continuously train the localization model and counteract the effect of signal variation. Specifically, a Hidden Markov Model (HMM) based Online Evaluation (HOE) method is firstly proposed to assess the accuracy of localization results by measuring the inconsistency of locations inferred by WiFi fingerprinting and Dead Reckoning (DR). Further, an Online Transfer Learning (OTL) algorithm is designed to improve the robustness of the fingerprinting localization, which consists of a weight allocation scheme to combine two classification models (i.e., the batch model and the online model) and an instance-based transferring scheme to resample the offline fingerprints and retrain the batch model. Finally, we implement the system prototype and give comprehensive performance evaluation, which demonstrates that the proposed solutions can outperform the state-of-the-art localization algorithms around 28% $∼$ 58% on vehicle tracking accuracy in dynamic indoor parking environments.

Publication
IEEE Transactions on Mobile Computing (中科院大类二区, CCF A 类期刊)
Ruitao Xie
Ruitao Xie
Associate Professor

My research interests include scheduling system resources to improve the performance of AI applications, edge computing, cloud computing, and mobile computing.