When vehicles are traveling on complex road surfaces, poor road conditions reduce driving safety and comfort. Road Surface Reconstruction (RSR) provides precise information about the road surface ahead of the vehicle, enabling predictive control of the magnetorheological adaptive suspension to enhance driving safety and comfort. RSR from Bird’s Eye View (BEV) has attracted much attention due to its potential for performance enhancement. However, existing methods for converting perspective views to BEVs face challenges such as information loss and representation sparsity. In this paper, we propose the BEV stereo model, which overcomes the limitations of the traditional viewpoint conversion module for pavement height reconstruction through BEV views. Deep learning is introduced in the image feature extraction module to improve the accuracy, and finally the three-dimensional height situation of the road in front of the vehicle is reconstructed on the dataset, which verifies the feasibility of the model and provides a technical way to improve the safety and comfort of the self-driving car. Compared with the existing model, the model in this paper has an absolute height error of 0.500cm and a root mean square error of 0.610cm, which is more feasible and superior.