基于GRU的可见光LOS/NLOS识别与高精度室内定位方法

GRU-Based Visible Light LOS/NLOS Identification and High Precision Indoor Positioning Method

摘要: 针对可见光室内定位系统中非视距(NLOS)传播导致定位精度急剧下降的问题，本文提出了一种基于门控循环单元(GRU)的LOS/NLOS识别与高精度定位方法。首先，构建了包含均值、标准差、峰度、偏度及中位数等5维统计特征的输入向量，以全面刻画接收信号强度(RSS)序列在视距与非视距条件下的分布差异与时序特性。其次，设计并训练了双层GRU深度学习网络，通过挖掘特征间的非线性依赖关系，实现了对链路状态的精准分类，输出各LED链路的LOS置信概率。在此基础上，提出了一种基于概率排序的动态LED优选策略，摒弃传统的硬判决剔除机制，转而选取LOS概率最高的三个LED作为定位锚点，并结合最小二乘法(LS)进行坐标解算，有效平衡了信号可靠性与几何精度因子。仿真结果表明，该方法在随机单遮挡场景下的LOS/NLOS识别准确率达到78.56%，LOS召回率高达82.95%；在定位性能方面，平均定位误差由基准方法的1.046 m降低至0.620 m，降幅达40.7%，且累积分布函数(CDF)曲线紧密逼近理论最优解(实际LOS)。该研究验证了所提方法在抑制NLOS误差、提升系统鲁棒性方面的显著优势，为复杂室内环境下的高精度可见光定位提供了有效的技术途径。

Abstract: Addressing the severe degradation of positioning accuracy caused by Non-Line-of-Sight (NLOS) propagation in Visible Light Positioning (VLP) systems, this paper proposes a high-precision positioning method based on Line-of-Sight (LOS)/NLOS identification using Gated Recurrent Units (GRU). First, a 5-dimensional statistical feature input vector, comprising mean, standard deviation, kurtosis, skewness, and median, is constructed to comprehensively characterize the distribution differences and temporal characteristics of Received Signal Strength (RSS) sequences under LOS and NLOS conditions. Second, a dual-layer GRU deep learning network is designed and trained to mine nonlinear dependencies among features, achieving precise classification of link states and outputting the LOS confidence probability for each LED link. Building on this, a dynamic LED selection strategy based on probability ranking is proposed. Departing from traditional hard-decision exclusion mechanisms, this strategy selects the three LEDs with the highest LOS probabilities as positioning anchors and employs the Least Squares (LS) method for coordinate estimation, effectively balancing signal reliability and Geometric Dilution of Precision (GDOP). Simulation results demonstrate that in random single-occlusion scenarios, the proposed method achieves an LOS/NLOS identification accuracy of 78.56% and an LOS recall rate of 82.95%. In terms of positioning performance, the average positioning error is reduced from 1.046 m in the baseline method to 0.620 m, representing a 40.7% improvement. Furthermore, the Cumulative Distribution Function (CDF) curve of the positioning error closely approximates the theoretical optimum (Ground Truth LOS). These findings validate the significant advantages of the proposed approach in mitigating NLOS errors and enhancing system robustness, offering an effective technical pathway for high-precision VLP in complex indoor environments.