<p>3D object detection represents a crucial component of robot perception. Current point cloud-based 3D object detection methods struggle to handle occluded objects effectively. In this paper, we propose two novel modules to improve the performance of 3D object detection on point clouds for occluded objects. Firstly, we introduce a novel post-processing module, semantic boundary clustering and processing (SBCP), which leverages semantic feature extraction and clustering of boundary points to robustly assess low-confidence proposals, thereby retaining bounding boxes for occluded objects effectively. Secondly, we develop a localized occlusion regulation approach that quantifies occlusion by analyzing the discrepancy between local resolution and point cloud density. This method dynamically adjusts the weights of the occluded targets during model training, mitigating the adverse effects of occlusion on detection accuracy. We evaluate our framework on the widely used KITTI dataset, as well as on our OTLD dataset, which presents greater challenges in terms of occlusion. Our method demonstrates performance on par with state-of-the-art approaches for non-occluded objects and significantly outperforms competing methods in detecting occluded objects. The ablation experiments demonstrate that the proposed modules significantly improve the model detection performance for occluded objects.</p>