Accurate and robust point cloud registration is essential for many robotic applications, but it heavily depends on the consistency of the environment. Dynamic objects in cluttered environments introduce noise, degrading registration performance. To mitigate it, various filtering approaches have been proposed, including geometric and deep learning-based methods. However, these methods may struggle with semi-static objects due to their reliance on object movement or the need for large labeled datasets, which are time-consuming to create. In this work, we propose a novel automated method for generating training data for dynamic object filtering in 3D LiDAR point clouds. Our approach integrates occupancy maps and long-term data acquisition to label and detect low-dynamic and semi-static objects without manual annotation. The labeled dataset is then used to train a semantic segmentation model, improving point cloud registration accuracy by removing dynamic objects before alignment. We validate our method using three real-world warehouse datasets, with a manually annotated subset, demonstrating the effectiveness of the method in improving Iterative Closest Point registration accuracy in robotic-like scenarios.