مقارنة محكومة بين النماذج الكلاسيكية والعصبية لتحديد المواقع باستخدام الضوء المرئي داخل المباني

الملخص

Indoor visible light positioning (VLP) studies frequently rely on simplified channel and noise models. In particular, optical signals emitted by multiple LEDs may be superimposed at each photodiode (PD), while receiver noise is sometimes approximated as a fixed proportion of the received signal. In this study, an idealised simulation framework is adopted in which individual LEDs are treated as distinguishable sources, representing systems employing time-, frequency-, or code-division multiplexing. Single-bounce non-line-of-sight (NLOS) reflections from the four vertical walls are incorporated using an adaptive mesh, while receiver noise is modelled through shot and thermal noise components.

Six learning-based positioning models are evaluated: a multilayer perceptron (MLP), a convolutional recurrent neural network (CRNN), a radial basis function support vector machine (RBF-SVM), a random forest, a k-nearest-neighbour (KNN) regressor, and a decision tree. Their performance is benchmarked against an Oracle maximum-likelihood reference and a Centre baseline across two room geometries, four receiver heights, and LED configurations ranging from one to sixteen sources. The RBF-SVM achieves the lowest mean root mean square error (RMSE) for (N_{\text{LED}}\leq 8), whereas the MLP performs best when sixteen LEDs are used. For (N_{\text{LED}}\geq 4), both models achieve mean RMSE values ranging from 0.10 to 0.22 m. The CRNN provides no improvement over the MLP under the evaluated conditions, while the tree-based and instance-based methods exhibit higher positioning errors.

Notably, in the (10\times10\times5) m room, the Oracle maintains a near-zero positioning error with sixteen LEDs. This finding suggests that the higher errors observed in the learning-based models may primarily arise from limitations in the learning framework rather than from insufficient positioning information in the simulated optical channel. These results highlight the importance of model selection and learning configuration when evaluating machine-learning-based VLP systems under idealised source-separation assumptions.