CRLB-based Data-driven Covariance Tuning for 5G KF Vehicular Tracking

Paper in proceeding, 2024

5G mmWave offers a high-precision positioning solution, functioning effectively in both line-of-sight (LoS) and operable non-line-of-sight (NLoS) conditions. However, in scenarios with complete signal blockage, integrating with motion-based models becomes crucial. This integration is achieved through Bayesian-based estimators, which entail a prediction and a correction stage weighted by their respective covariance matrices. Although covariance matrices of different prediction models have been extensively studied in the literature, the measurement covariance matrix derived from 5G-based position computations remains largely unexplored. In this paper, we propose a measurement covariance matrix tuning scheme based on a data-driven Cramér-Rao lower bound CRLB model. We validate the proposed algorithm within a simple linear Kalman filter (LKF) positioning framework. The methodology was tested in a controlled simulation scenario using a real 24-minute-long vehicular trajectory in a deep-urban environment. The results demonstrate that the developed data-driven model is reliable, maintaining a standard deviation error of less than 7 cm for 95% of the time and less than 0.5 m for 100% of the time relative to the true computed CRLB. The proposed adaptive KF sustains a position error below 30 cm for 99.3% of the time.