Neural Error Compensation in PPP-RTK for Vision-Inertial Kinematic Positioning

Abstract

Precise Point Positioning with Regional augmentation (PPP-RTK) provides absolute positioning at the decimetre-to-centimetre level without baseline constraints, yet residual tropospheric, ionospheric, and multipath errors continue to limit convergence performance in kinematic scenarios. Vision-Inertial Odometry (VIO) is drift-resilient over short intervals but accumulates unbounded
error without an external position anchor. This paper presents a tightly coupled neural error-compensation framework that integrates a Long Short-Term Memory (LSTM) network into a PPP-RTK/VIO fusion pipeline to predict and correct dominant GNSS residual errors in real time. The LSTM consumes carrier-phase observables, satellite geometry metrics, and inertial pre-integration
residuals, enabling environment-aware compensation without explicit atmospheric model inversion. A factor graph back-end unifies PPP-RTK pseudorange and carrier-phase factors with VIO pre-integration constraints and neural correction factors under an iSAM2 incremental smoother. On a mixed urban–suburban kinematic dataset, the proposed pipeline yields a modest but consistent
reduction in horizontal RMS positioning error (approximately 25.5%, from 4.7 cm to 3.5 cm) and in time-to-convergence (approximately 18.3%, from 18.3 min to 15.0 min) relative to a conventional PPP-RTK/VIO baseline without neural augmentation. These preliminary, single-route results indicate that learning-based residual compensation may be a useful, low-overhead supplement to kinematic GNSS positioning in moderately degraded signal environments; broader validation across routes, seasons, and receiver classes is needed before stronger claims can be supported.