Over the past few years, there has been a growing interest in indoor positioning systems that use Visible Light Communication (VLC) technology in conjunction with light-emitting diodes (LEDs). These systems have become popular because of their ability to provide high bandwidth, and it is anticipated that wireless communication may expand into the visible-light spectrum in the future. This makes VLC a promising option for high-speed line-of-sight communication. Additionally, the visible light spectrum can be deployed in the Industrial Internet of Things environment, as it does not emit electromagnetic radiation and is immune to electromagnetic interference. This article presents a database consisting of an average of 356 image samples obtained from a CMOS sensor. This database includes seven different classes, each class varying its frequency from 1 kHz to 4.5 kHz with a 500 Hz interval. The goal is to apply this database to various neural networks based on Extreme Learning Machines (ELM) in its variants: (1) Standard ELM, (2) Regularized ELM, and (3) Weighted ELM. MATLAB environment was used to evaluate the performance of the proposed indoor positioning system enabled for VLP. Our results indicate that a promising option is standard ELM since it has a performance greater than 99% accuracy and mean G, and low computational complexity compared to CNN.