This paper tackles indoor receiver localization as a regression task using visible light positioning (VLP). This technology has gained prominence as an alternative to conventional indoor positioning. Traditional systems such as WiFi and ZigBee face notable drawbacks: electromagnetic interference, added infrastructure, limited accuracy, and high deployment costs. In general, classical positioning approaches—lateration, triangulation, proximity, and least squares—typically infer location from measurements like received signal strength (RSS), time of arrival (TOA), time difference of arrival (TDOA), or angle of arrival (AOA). However, in real-world VLP scenarios, these methods face difficulties since they depend on ideal signals, which means unaffected by physical obstacles and perfectly aligned between transmitters and receiver. Our work proposes the application of the extreme learning machine (ELM) algorithm, its regularized version (R-ELM) and the Levenberg-Marquardt algorithm (LMA), to predict the receiver’s position based simply on the light signals emitted by light-emitting diodes (LEDs) and captured by a photodiode (PD). To evaluate the overall performance of these models, two experimental datasets were used: one in a 2D environment with 7344 RSS samples and another novel in a 3D environment with 27000 training records and 18000 testing samples by analyzing power spectral densities. To assess performance and analyze parameter optimization of ELM, R-ELM, and LMA, metrics such as root mean square error (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RMSE</i>), correlation coefficient (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i>), training and testing times were used. By considering the state of the art, the LMA model proved to be the most accurate in the 2D environment (RMSE less than 10 cm), although its training time was excessive (over one hour), incompatible with for real-time applications (time variant channels). In contrast, the R-ELM achieved an accuracy of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RMSE</i> = 12.4806 cm and completed its training in 31.2 ms (real-time learning), positioning it as a much more efficient alternative compared to LMA. For the 3D dataset, the R-ELM also showed the best overall performance, with an average RMSE close to 22.4784 cm and a training time of 1.0250 s. Incorporating both the dominant frequency and power as inputs into the ELM enhanced the model’s accuracy compared to using each input separately. Lastly, RMSE was evaluated as a function of the distance from the center of the room of each data set, concluding that the R-ELM is more stable in areas with weak signals or missing data, which is particularly valuable in real environments where the signal is not always homogeneous. Finally, the results demonstrate that it is possible to achieve a precise and fast positioning system based on VLP using the R-ELM, as this model generally achieves higher accuracy and computational efficiency.
10
