Terrain-Aware Adaptive Multisensor Fusion for Millimeter-Level Localization on Undulating Surfaces

Abstract

This study presents a terrain-aware adaptive multisensor fusion framework for high-precision localization of robotic platforms operating on slightly undulating terrain. The framework integrates heterogeneous sensing modalities, including RGB-D perception, inertial measurements, wheel odometry, and total-station observations, to perform online uncertainty-aware state estimation. Under total-station-constrained global localization on slightly undulating terrain, the laser-tracker-validated proposed framework maintains over 90% of localization errors within approximately 2 mm, restricts angular error within 0.2°, reducing the 90% cumulative distribution function (cdf) positional error from approximately 9–10 mm to around 2 mm relative to the conventional error-state extended Kalman filter (ES-EKF) framework. With an average latency of 22 ms/cycle, this lightweight framework supports real-time execution and mitigates high-frequency localization fluctuations caused by terrain-induced mechanical shocks. It should be noted that the reported millimeter-level accuracy relies on global observations from the total station, while the onboard heterogeneous sensors primarily serve to mitigate high-frequency terrain disturbances.