Zero-cost and Map-Free Shop-Level Localization Algorithm Based on Crowdsourcing Fingerprints.

Jie Wei,Xiaoyun Zhou,Fang Zhao,Haiyong Luo,Langlang Ye

DOI: https://doi.org/10.1109/upinlbs.2018.8559708

2018-01-01

Abstract:In recent years, with the rapid development of mobile Internet, intelligent location-based services attract widespread attention. Using the situated shop location of users, the accuracy of the message and advertisement delivery can be effectively promoted. Though the Wi-Fi based positioning has become the mainstream of indoor positioning technology, the large deployment and maintenance cost limits its widespread use. In this paper, we proposed a novel zero-cost and map-free shop-level localization algorithm. The algorithm collected Wi-Fi fingerprinting data and opportunistic GPS data when customers paid the bill through mobile payment app. This algorithm takes each shop identification as a location index and localizes user without the need of the map. The Wi-Fi fingerprints and GPS data are collected in the background when crowdsourcing users employ the mobile payment APP, so our proposed localization algorithm can work without any site survey. To improve the positioning accuracy and speed up localization calculation, we introduce a hierarchical classification method based on ensemble learning model. The upper classifier dynamically selects several local candidate shops with an online Wi-Fi fingerprinting, with the virtue of reducing the computational complexity. After completing the upper classification, we further perform a confidence-weighted ensemble classification combing using GBDT, random forest and KNN to estimate the user location. By aggregating the location estimations of individual classifiers with their corresponding positioning accuracy, the confidence-weighted ensemble classification model can obtain more accurate location estimation. The paper also introduces some features, i.e., historical shop scanning Wi-Fi ratio, Wi-Fi signal statistics, store type and transaction time distribution, to improve the accuracy of the bottom shop classifier. Extensive experimental results show that our proposed method outperforms the maximum likelihood estimation and KNN algorithm, and can localize user with more than 92% accuracy.