In visual SLAM systems for autonomous vehicles, the influence of feature distribution leads to fewer vertical constraints, resulting in significant vertical drift and affecting the long-term localization performance of the system. This paper proposes a graph optimization-based stereo visual SLAM system with road constraints, which enhances vertical constraints by extracting more features from the road and establishing explicit constraints between the vehicle and the road plane. First, a stereo matching method for road feature points is proposed to compensate for the disparity-induced road feature offset between the left and right images, improving the system’s ability to extract road features and enhancing feature distribution. Then, the local road plane is used to represent the road, and explicit constraints between the road and the vehicle are established to further increase vertical constraints on the system. Finally, the local road plane, vehicle pose, and map points are optimized jointly as nodes in a nonlinear optimization. Validation through the KITTI dataset and real vehicle experiments shows that the proposed system reduces vertical drift and achieves more accurate localization results.