Deep Learning Based Fingerprints Reduction Approach for Visible Light-Based Indoor Positioning System

摘要

Received signal strength and fingerprints based indoor positioning algorithm has been commonly used in recent studies. The actual implementation of this method is, however, quite time-consuming and may not be possible in large spaces, mainly because a large number of fingerprints should be collected to maintain high positioning accuracy. In this work, we first propose the deep learning-based fingerprints reduction approach to reduce the data collection workload in the offline mode while ensuring low positioning error. After estimating the extra fingerprints using a deep learning model, these new fingerprints combine with the initially collected fingerprints to create the whole training dataset for the real estimation process. In the online mode, the final estimated location is determined using the combination of trilateration and k-nearest neighbors. The experiment results showed that mean positioning errors of 1.21 cm, 6.86 cm, and 7.51 cm are achieved in the center area, the edge area, and the corner area, respectively.