Multi-Directional Light-Emitting Diodes Aided Visible Light Positioning Scheme Based on Neural Network
Objective The exponential growth of indoor mobile traffic and data connections has led to an increasing demand for indoor positioning.Since satellite-based positioning systems like GPS and Beidou face significant limitations indoors due to signal attenuation and blockage,it is crucial to find a high-accuracy,low-cost indoor positioning solution.Visible light positioning(VLP)has emerged as a strong candidate for indoor positioning due to its low cost,high accuracy,lack of electromagnetic interference,and ability to simultaneously support communication,positioning,and illumination.Currently,VLP technologies based on photodetectors(PDs)have been widely studied and applied because of their low cost and ease of implementation.These systems typically estimate user coordinates through various positioning algorithms,such as time difference of arrival(TDOA),angle of arrival(AOA),received signal strength(RSS),and RSS fingerprinting with artificial neural networks(ANNs).However,these methods often assume that light sources are vertically aligned and the PD is positioned horizontally.In dynamic scenarios where the PD is randomly tilted,positioning can degrade.To address this issue,researchers have proposed a hemispherical lamp with multi-directional LEDs(MD-LEDs)and a positioning algorithm based on the RSS ratio(RSSR).While this method can accommodate tilted PDs,the hemispherical lamp structure is difficult to implement,and the algorithm details are not fully disclosed.To overcome these limitations,other researchers have developed the first prototype of a multi-directional LED integrated lamp,introducing the RSSR-based linear least squares(RSSR-LLS)algorithm.Experimental results show that the average localization error(LE)of the RSSR-LLS algorithm is about 11 cm when the PD is tilted in a relatively open environment.To further enhance the positioning performance of the MD-LED-aided VLP(MD-VLP)system,we propose a new approach using a neural network(RSSR-ANN)based on the RSSR method.Extensive simulations and experiments confirm that this new method achieves the desired performance improvements.Methods In this study,we propose an RSSR-ANN scheme for the MD-VLP system.This approach utilizes a single PD to capture light intensities from multiple LEDs oriented in different directions,then feeds the RSSR values between pairs of LED intensities into a trained artificial neural network model to predict plane coordinates.Specifically,a feed-forward backpropagation neural network model with three hidden layers is constructed,including an input layer,hidden layers,and an output layer(Fig.2).The RSSR values from known positions are used as input,and the corresponding plane coordinates serve as output.Model training is performed using gradient descent and backpropagation to determine the weights and biases,resulting in the final target model.Once the ANN model is established,it can predict the user's plane coordinates by processing the RSSR values received by the PD.The neural network can automatically extract features from vast amounts of data and adjust the network weights iteratively through its multi-layer hidden structure,effectively fitting nonlinear data.Therefore,the method also supports MD-VLP systems using non-Lambertian radiation LEDs.Results and Discussions To verify the universality of the proposed scheme,we simulate and analyze the positioning performance of the MD-VLP system using LEDs with different radiation patterns(Fig.4).MATLAB software is used to simulate PDs receiving signals from multi-directional LEDs,generating RSSR values at various coordinates.These values are then fed into a Pytorch®-based ANN model for training.Once trained,test data are used to evaluate the model's performance.Simulation results demonstrate that when using two types of non-Lambertian radiation LEDs,the average LE achieved is 3.70 cm and 6.22 cm,respectively,even with moderate PD tilting(Fig.5).In addition,we conduct a series of experiments using both MD-LED integrated lamps with standard Lambertian LEDs and a VLP device consisting of commercially available LED downlights to validate the generalizability of the proposed scheme(Fig.6).Experimental results indicate that the average LE is 2.96 cm when the PD is horizontal and 5.51 cm when tilted at 15°(Fig.8),a significant improvement over the RSSR-LLS algorithm.Furthermore,we investigate the positioning performance of the RSSR-ANN scheme in a random reflection environment,where reflective objects like tables and chairs are arranged around the PD to alter the illuminance distribution.Results show that the scheme achieves an average LE of 6.02 cm even when the PD is horizontal.Conclusions Several key conclusions can be drawn from the simulation results.First,the proposed RSSR-ANN scheme is suitable for MD-LED integrated lamps using standard Lambertian radiation LEDs as well as non-standard Lambertian LEDs.Second,the MD-VLP system can utilize non-Lambertian radiation LEDs,as long as the radiation curve remains monotonic within the angle range[0°,θm],where θm is the maximum radiation angle.Finally,the ANN model trained with a horizontal PD still supports positioning when the PD is randomly tilted.Compared to the RSSR-LLS algorithm,the RSSR-ANN scheme increases average LE by 40%.Experimental verification further demonstrates that the proposed scheme is effective in systems using both standard and non-standard Lambertian LEDs,enhancing the overall universality of the MD-VLP model.In addition,the RSSR-ANN scheme performs well in dynamic scenarios with randomly tilting PDs,improving LE performance by 52.5%-72.3%under various conditions(Table 4).This suggests that the MD-VLP system based on RSSR-ANN holds great potential for mobile applications.
received signal strength ratioartificial neural networkmulti-directional light-emitting diodevisible light positioning
NETL
NSTL
万方数据
