Reducing the burden of data collection in a fingerprinting-based VLP system using a hybrid of improved co-training semi-supervised regression and adaptive boosting algorithms

Abstract

The positioning accuracy of the visible light-based fingerprinting technique depends greatly on the number of initially collected and labeled data points. The greater is the number of initial fingerprints collected, the higher is the achieved accuracy of the system. However, collecting many labeled fingerprints while simultaneously ensuring data reliability is quite time consuming. To increase the practicality and applicability of fingerprinting-based algorithms, we combine the improved co-training semi-supervised regression and adaptive boosting (Ada-XCoReg) algorithms. Ada-XCoReg is a combination of double-weighted k-nearest neighbors, cross-correlation, and an adaptive boosting technique that increases the amount of input data and enhances positioning accuracy based on randomly unlabeled data points. By applying the proposed solution to the experimental model, we proved that our approach achieved a positioning accuracy of 6.14 cm, 5.29 cm, and 2.51 cm when the numbers of initially-labeled fingerprints were 16, 25, and 49, respectively.