ReCPos: Deep Learning Network for 5G NR High-precision Positioning

Abstract

With the increasing demand for location-based services, various positioning technologies have been proposed. However, the existing positioning technologies are often limited to the line-of-sight (LOS) scenario and incapable of providing accurate results under non-LoS (NLOS) scenario. Therefore, we propose ReCPos net, a deep residual convolutional neural network for high-precision positioning in heavy NLOS scenario, where four residual modules with distinct structures are designed. After the preprocessing for the channel impulse response (CIR) and reference signal received power (RSRP), the designed residual modules are adopted to extract the high-dimensional feature vector, then the location can be predicted by the ReCPos net. The experiment results indicate that the proposed ReCPos could reduce the positioning error by at least 20.0% compared to the existing AI-based positioning schemes in heavy NLOS conditions with lower model complexity and computing power. The input of CIR combined with RSRP and truncating CIR outperforms that of CIR alone in terms of positioning accuracy.