基于改进ShuffleNet V2的室内可见光指纹定位算法

Indoor Visible Light Fingerprint Positioning Scheme Based on Improved ShuffleNet V2

为增强室内可见光定位系统准确性及实时性,提出了基于改进ShuffleNet V2神经网络的可见光指纹定位算法。该算法使用参考节点的接收信号强度和到达角度作为联合指纹特征,并使用接收器坐标作为训练标签来创建指纹数据库。通过消融实验,改进了ShuffleNet V2网络结构,将指纹数据库引入改进ShuffleNet V2网络训练,实现不同数据通道之间的信息交流,提高了神经网络的特征提取能力。为比较和分析其他多种可见光室内定位算法的性能,体现所提算法的技术优势,进行了仿真实验,在室内5 m×5 m×3 m的定位场景下,提出的定位算法平均定位时间为174 ms,平均定位误差为2.30 cm,同时针对室内不同高度的定位精度也达到厘米级。

Objective In recent years, with the rapid development of wireless sensor networks and Internet of Things(IoT) technology, indoor positioning technology has been widely used in several fields, such as robot navigation. How to achieve accurate and fast positioning on the mobile platform has become a hot topic of discussion in the academic community. Visible light positioning technology based on visible light communication has the advantages of wide spectrum, good security and low cost, which makes it one of the most promising visible light positioning technologies. Common positioning methods for visible light positioning include received signal strength(RSS), time of arrival(TOA), time difference of arrival(TDOA), and angle of arrival(AOA). Neural network is a common technical means for indoor visible light positioning, but its shortcomings such as large quantity of parameters and long computing time limit the deployment of mobile platform. The emergence of high-performance and lightweight neural networks such as MobileNet and ShuffleNet provides the conditions for the deployment of visible light positioning mobile platform. In order to enhance the accuracy and real-time performance of the indoor visible light positioning system, a visible light fingerprint positioning algorithm based on improved ShuffleNet V2 neural network is proposed.Methods In a 5 m×5 m×3 m indoor scene, four transmitter light-emitting diodes(LEDs) are uniformly distributed on the ceiling,and the receiver is located on the ground at a height of 0 m. The channel model for line-of-sight transmission of indoor visible light communication is established, and the measurement formulas of RSS and AOA are given. The ground is divided into 2601 small cells of 0.1 m×0.1 m. The RSS and AOA at the center of each cell are collected separately, and the corresponding coordinates are recorded to construct a fingerprint library. In order to determine the suitable network structure, ablation experiments are conducted on the ShuffleNet V2 network, and the positioning network structure(Fig. 12) is finally determined to compare the effects of different optimizers and learning rates on the positioning performance. The data from the fingerprint library are normalized and inputted into the improved ShuffleNet V2 network for training, preserving the trained model fixed. In order to test the generalization of the positioning algorithm, the training set is located at the ground level and fixed, and the test set is collected in the plane at heights of 0.5 m, 1 m and 1.5 m for positioning. High accuracy positioning results are obtained.Results and Discussions Simulation experiments are performed based on the indoor channel model. 2601 points are uniformly selected on the ground at a height of 0 m as the test set, and the data are inputted into the trained improved ShuffleNet V2 network for positioning(Fig. 13). The positioning error fluctuates between 0.26 cm and 9.25 cm, with an average positioning error of 2.30 cm and an average positioning time of 174 ms. The positioning error is larger near the light source and indoor edges. Fixing the simulation parameters and comparing the positioning performance of the proposed method, convolutional neural network(CNN) and backpropagation neural network(BPNN), the proposed model converges around 10 epochs and the training set is well fitted to the validation set(Fig. 15). Five repeated experiments are conducted respectively, showing that the error of the proposed model fluctuates the least(Fig. 16). The cumulative distribution function(CDF) curves of the proposed model are shifted left(Fig. 17)compared with those of CNN and BPNN, and are within the 95% confidence interval. The positioning errors of the three algorithms are 4.93 cm, 9.25 cm and 17.77 cm, respectively. Compared with single RSS fingerprint, the average positioning accuracy of this algorithm is improved by 54.1%, which indicates the superiority of the joint fingerprint feature positioning algorithm;in the case of unchanged fingerprints, compared with the CNN algorithm, the average positioning accuracy of the present algorithm is improved by 43.8%(Table 4). In the case of different heights(0.5 m, 1 m, 1.5 m) of the test set, the training set is still located on the ground and 121 groups of test samples are collected at intervals of 0.5 m under each height for the positioning experiments. For the height h=0.5 m, the average positioning error is 5.44 cm;for h=1 m, the average positioning error is 8.44 cm;and for h=1.5 m, the average positioning error is 16.31 cm.Conclusions To enhance the accuracy and real-time performance of indoor visible light positioning system, a visible light fingerprint positioning method based on improved ShuffleNet V2 is proposed. The method uses RSS and AOA of the reference nodes as joint fingerprint features and the receiver coordinates as training labels to construct the fingerprint database. Through ablation experiments, the ShuffleNet V2 network structure is improved, and the fingerprint database is introduced into the improved ShuffleNet V2 network training to achieve information exchange between different data channels and improve the feature extraction capability of the neural network. Simulation experiments are conducted for comparing and analyzing the performance of various visible light indoor positioning algorithms and confirming the technical advantages of the algorithm proposed in this paper. Under the indoor 5 m×5 m×3 m positioning scenario, the algorithm has an average error of 2.30 cm and an average positioning time of 174 ms. It proves the feasibility and efficiency of the proposed positioning method. Even if the fingerprint database data and the positioning data are located in planes at different heights, the centimeter-level positioning accuracy can be achieved, which meets the needs of most indoor positioning services.