In recent years, the digital preservation of immovable cultural heritage has increasingly relied on 3D reconstruction technologies, with 3D Gaussian Splatting (3DGS) favored for its high-efficiency and high-fidelity rendering quality. However, due to the challenges of performing systematic and controlled image capture at immovable heritage sites, crowd-sourced data offer a low-cost alternative with broad, diverse coverage, albeit often affected by occlusions that degrade 3D reconstruction quality. To address this challenge, we propose an occlusion-aware 3D reconstruction pipeline based on 3DGS and crowd-sourced image collections. Our method integrates an object detection module to identify dynamic occlusions and generate binary masks, followed by a LaMa-based image inpainting model to semantically restore occluded regions. These refined images effectively enhance point cloud quality and are used for high-fidelity reconstruction. Additionally, we incorporate DINOv2 features for pixel-level refinement, improving geometric consistency and the overall accuracy of the reconstructed 3D models. Extensive experiments conducted on multiple public datasets containing immovable cultural heritage demonstrate that our method achieves superior performance in both occlusion removal and reconstruction quality compared to existing approaches, highlighting its strong generalization capability and practical value.