Research on road surface perception strategy for intelligent suspension based on vision and dynamics fusion

Abstract Road conditions in real-world driving are highly complex and subject to frequent changes. Traditional perception strategies often struggle to simultaneously achieve accurate road type classification and real-time estimation of roughness grade, especially under aging infrastructure, where two key challenges arise: the non-uniqueness of the ‘road types and roughness grades’ mapping, and the difficulty in accurately estimating impulsive feature parameters. To address these issues, this paper proposes a multimodal fusion road perception strategy for intelligent suspension systems, aiming to enhance perception accuracy and robustness by integrating visual imagery and dynamic response data. First, a data acquisition platform is developed to collect typical datasets covering both random and impulsive road surfaces. To resolve the non-uniqueness of the ‘road types and roughness grades’ mapping, a convolutional block attention module deeplabV3(CBAM-DeepLabV3) semantic segmentation network is used to extract visual road type features, and a random forest extreme gradient boosting (RF-XGBoost) model is employed to estimate road roughness grades. This fusion approach reveals the inherent ambiguity in the mapping between road types and roughness levels. A dynamic mapping inference mechanism is then constructed to enable adaptive updates of the mapping rules. To address the difficulty in estimating impulsive feature parameters, a novel method combining inverse perspective transformation with reinforcement learning is proposed. This approach enables multidimensional measurement of impulsive features, including longitudinal distance, width, and vertical height, significantly improving both parameter recognition accuracy and real-time performance. The proposed method is validated using a co-simulation platform based on Prescan, Carsim and Simulink. Test results show that the method achieves high recognition accuracy and strong generalization capability in road perception tasks, providing robust prior perception support for intelligent suspension control systems.