Robust GNSS/INS navigation using dynamic kernel weighting under abnormal observations

Abstract

This work addresses robust GNSS/INS navigation in the presence of abnormal observations. To mitigate the impact of GNSS outliers and time-varying measurement disturbances, an innovation residual-driven dynamic kernel weighting scheme is incorporated into the error-state extended Kalman filter. Several robust kernels (L2, Huber, Cauchy) are evaluated, and an online measurement noise adaptation strategy is further adopted to enhance filter update stability. Experiments are performed on real vehicle data from the KITTI dataset, where multiple anomalous GNSS conditions — including burst outliers, signal outages, multipath interference, and time-varying noise — are artificially injected. The results indicate that the proposed dynamic weighting approach can effectively attenuate the influence of degraded measurements and yield improved localization robustness compared with conventional robust filters. The method also maintains low computational cost, resulting in better trajectory accuracy and position error performance under complex disturbance scenarios.