Special vehicles such as off-road vehicles and planetary rovers frequently operate on complex, unpaved road surfaces with varying mechanical parameters. Inaccurate estimation of these parameters can cause subsidence or rollover. Existing methods either lack proactive perception or high precision. This article proposes a fusion framework integrating a visual classifier and a dynamics observer for stable, accurate estimation of road surface parameters. The visual classifier uses an adaptive segmentation system for unpaved roads, leveraging a large-scale vision model and a lightweight network to classify upcoming road surfaces. The dynamics observer employs an online wheel-–ground interaction model using stress approximation, integrating strong tracking theory into an unscented Kalman filter for real-time parameter estimation. The fusion framework performs integration of the classifier and observer outputs at data, feature, and decision levels. An adaptive fading factor and recursive Gaussian process modeling ensure precise estimation of varying parameters. Real-vehicle tests demonstrate that the proposed method reduces the average estimation error by 8.5% and improves convergence speed by 40% during road surface changes, demonstrating potential for integration into off-road vehicle stability control systems.
Zhang et al. (Fri,) studied this question.