Accurate online surround-view system calibration is crucial for autonomous driving, as even small sensor shifts or sudden vibrations can lead to significant extrinsic deviations and compromise the reliability of 360 \(^{\circ }\) environment perception. Existing methods often rely on single-modality information and are highly sensitive to image distortions or initial parameter settings, limiting their applicability to diverse real-world scenarios. We introduce OSCalib, a novel end-to-end Online Surround-view Calibration framework for surround-view systems that combines information from two sensing modalities with a MAP-based initialization scheme to tackle these issues. Specifically, our approach integrates ground semantics–which offer stable and metrically rich references–with SLAM-based motion estimation to establish the GSAlign (Ground-Surround Alignment) model, significantly reducing the feature instability caused by image distortions. Furthermore, OSCalib introduces an MAP-based system initialization strategy to jointly solve for SLAM system scale and motion transformations, ensuring a robust startup configuration even under substantial sensor perturbations. Finally, to address occlusion-induced tracking failures in motion estimation, we propose a robust uncertainty-aware semantic data association strategy. It addresses the limitations of existing methods in handling uncertainty by adaptively adjusting geometric covariances of ground semantics. Extensive quantitative and qualitative experiments across diverse real-world scenarios, including indoor and outdoor environments with varying road textures, demonstrate its superior accuracy and generalization ability for online surround-view system calibration.