Efficient Deep Learning-Based mmWave Positioning Through Ray-Tracing Method

Deep learning-based positioning methods, known for their ability to process rich scatter signals in complex scenarios, present advantages over conventional methods. However, to achieve accurate positioning for the neural network, a large amount of training data is required, which is very time-consuming and high cost to collect. To address this, our work leverages ray-tracing techniques for efficient channel modeling and real dataset collection. In this paper, we present a practical and scalable alternative, a ray-tracing-driven data generation pipeline combined with a deep-CNN estimator and an SNR-aware multi-resolution grid refinement stage. By utilizing a 3D model of the actual environment, we adopt the ray-tracing method to accurately calculate path information, which is then used to model millimeter-wave (mmWave) multiple-input multiple-output (MIMO) channels and generate received signals for user equipment (UEs) as training data. This efficient approach eliminates the need for physical relocation of UEs and extensive on-site data collection, significantly enhancing the practicality and applicability of deep learning-based positioning techniques. Additionally, we refine the neural network's coarse positioning results with a grid search technique to further enhance accuracy. Simulation results show that the proposed ray-tracing-based method achieves sub-meter-level positioning precision, even in changing environments, validating the effectiveness of the proposed ray-tracing method for data generation.