Objective With the rapid development of the Internet of Things industry, the demand for indoor positioning solutions is increasing. Visible light indoor positioning technology has attracted increasing attention in research and application development. Currently, LED indoor positioning technology is still in its infancy. However, with the development of visible light communication technology, the visible light indoor positioning industry has also developed rapidly in recent years. Visible light indoor positioning is a new positioning technology that combines lighting and communication. Compared with traditional indoor wireless positioning methods, it has the advantages of low cost, no electromagnetic interference, high positioning accuracy, and broad prospects. However, existing visible light positioning technology has difficulty in handling background noise interference and indoor reflection noise, which leads to unstable positioning accuracy. An artificial neural network (ANN) is capable of nonlinear mapping, self-learning, self-adaptation, and generalization. Additionally, ANNs can extract key information from a large amount of data. ANNs have been applied to outdoor positioning based on the Global System for Mobile Communication. To reduce the interference of diffuse reflection of wireless optical channels to RSS-based visible light positioning systems and improve positioning accuracy, this paper proposes a high-precision indoor positioning algorithm based on a multiple reflection channel model and a neural network.
Methods This paper investigates visible light indoor positioning algorithms commonly used at home and internationally. Indoor positioning technology based on high-precision photoelectric sensors and image sensor imaging are compared and analyzed. The visible light indoor location algorithm based on a neural network is summarized and proposed to improve the location accuracy. First, a visible light indoor stereo positioning system is modeled. To avoid the influence of diffuse reflection of the light channel on positioning accuracy, a channel mathematical model that includes a direct line-of-sight link and first-order reflection link is established. Second, to achieve 3D positioning, the optical intensity data should be obtained by decoding different LEDS at different positions on different planes of different heights. These data can be used to create fingerprint databases. After determining the LED light source and channel model, combined with an LED channel diffuse reflection model, grid calibration is carried out on the receiving plane and the illumination intensity of different LEDs at the center point of each grid is collected. After classifying the collected data, the training and test datasets can be created. Third, the localization algorithm based on a BP (Back Propagation) neural network is designed, and data training and prediction is performed. The neural network positioning system designed in this paper is divided into input, hidden, and output layers. The input to the neural network is RSS from different LEDs. A BP neural network is used to fit the parameters of a real indoor wireless channel. The output of the neural network is an m-dimensional space vector used for coarse positioning of the target to be measured. This coarse positioning represents the relative spatial position of the position coordinates of the predicted receiver and the position coordinates of the LED light source. Finally, through the positioning error constraint model based on location variance and Euclidean distance to solve the positioning equation, the predicted position coordinates of the target to be tested is determined.
Results and Discussions To verify the theoretical validity and positioning reliability of the proposed algorithm, a simulated three-dimensional (3D) positioning experiment and an actual positioning experiment are performed. During the simulation positioning test, the 3D space with 4 m×4 m×3 m is taken as the model in this paper, and the simulation experiment is carried out according to the simulation parameters shown in Table 1. In the experiment, a training set with a spacing of 5 cm is divided on each plane, and the area of each rectangle is 5 cm × 5 cm. A reference fingerprint point is selected from each small square, and a total of 6561 points are selected from each plane. According to the results (Fig.4), when the plane with a height of 0.5 m is tested, there is no great deviation between the predicted position and the actual position coordinates. The maximum error is 7.10 cm and the average error is 1.73 cm; 90.1% of the error is within 3 cm (Fig.7). In the test plane with a height of 1.0 m, the maximum positioning error is 5.56 cm and the average error is 1.29 cm (Fig.5); 91.7% of them have positioning errors less than 3 cm (Fig.8). The maximum positioning error is 12.38 cm and the average error is 3.85 cm when the height plane is 1.5 m (Fig.6), and positioning error data within 3 cm accounts for 41.7% (Fig.9). In the measurement and positioning stage, experimental platforms are built in 0.8 m-long, 0.8 m-wide, and 0.8 m-high three-dimensional spaces (Fig.10). After multiple positioning tests are conducted on 81 groups of training data and 16 groups of position data, the average positioning error is 3.65 cm (Fig.12).
Conclusions Traditional visible light indoor positioning systems based on RSS are vulnerable to background noise and indoor reflection noise. In this paper, a visible light indoor positioning model based on a neural network is proposed by combining the neural network with RSS positioning technology. Combined with multiple reflection channel modeling, the model uses an ANN to study the visible light channel parameters to facilitate neural network training and model testing. The positioning error constraint model based on location variance and Euclidean distance is used for effective correction. These can significantly improve the indoor positioning accuracy of a visible light system to achieve accurate positioning.
