@article{Liu2026, 
author = {Jiahui Liu and Yang Liu and Chenglin Liu and Liang Wang and Xiaohan Liu},
title = {Predictive estimation of tire-road friction for autonomous driving via vision-dynamics collaborative fusion},
year = {2026},
journal = {Tsinghua Science and Technology},
keywords = {Autonomous driving, Tire-road friction estimation, Information fusion, Vehicle dynamics, Road surface condition classification},
url = {https://www.sciopen.com/article/10.26599/TST.2026.9010105},
doi = {10.26599/TST.2026.9010105},
abstract = {The tire-road friction coefficient (TRFC) is a decisive factor governing vehicle safety, yet estimating it accurately in real time remains a challenge for autonomous driving. Existing dynamics-based methods rely on vehicle dynamics responses and require sufficient tire excitation, while vision-based methods can preview upcoming road conditions but typically yield only coarse friction ranges rather than precise values. To fill this gap, this paper proposes a vision-dynamics collaborative fusion framework for predictive TRFC estimation. A road surface condition classifier augmented with a confusion-aware training strategy is developed to distinguish ”wet” from ”water” surfaces. The classification results, together with the ego vehicle’s kinematic state, are mapped into a continuously updatable preview friction field (PFF) that spatially anticipates the road friction ahead. The PFF is then adaptively fused with an adaptive cubature Kalman filter (A-CKF) estimator through a dual-confidence fusion mechanism. Real-vehicle experiments demonstrate that the proposed Fusion method consistently outperforms purely dynamics-based or vision-based baselines, achieving a mean absolute percent-age error as low as 4.48% in low-excitation conditions, where the dynamics-only A-CKF degrades to 42.60%. Moreover, benefiting from the PFF’s predictive capability, the Fusion method converges to the true TRFC markedly faster than the dynamics-only estimation. These results confirm the accuracy, robustness, and predictive capability of the proposed framework for autonomous driving applications.}
}