Multivariable Fingerprints With Random Forest Variable Selection for Indoor Positioning System

摘要

Indoor positioning technology with fingerprints based on 5th Generation (5G) mobile communication system has attracted extensive attention. Due to instability caused by non-line-of-sight and multipath propagation, received signal strength indicator (RSSI) values at different points may be similar. To address the problem and meet the requirement of high-precision indoor positioning, multivariable fingerprints with random forest variable selection (RFVS) is proposed in this paper. Ten 5G variables related to distance-power relationship are chosen to establish the multivariable fingerprinting database. In order to utilize correlation among the multivariable fingerprints, RFVS is used to sort variable importance and combinations. Five machine learning algorithms are used to calculate user equipment's positions in three experimental scenarios for test. Results show that multi-layer perception (MLP) achieves best performance among the five positioning algorithms in all test scenarios. Combined with RFVS, MLP shows 31.42%, 39.56% and 30.54% accuracy improvement compared with that of RSSI only used in three test scenarios respectively. Besides, the best result of MLP achieves 1.198 m positioning error for 80% test samples in one of three experimental scenarios.

关键词