<p>High precision 6-DoF pose estimation is a challenging and critical research topic for robot grasping, with existing models still facing significant challenges. First, objects may be stacked or occluded in complex scenes, resulting in incomplete data and invisible features, which significantly degrade the accuracy of pose estimation. Second, existing models trained on synthetic data may exhibit poor generalization ability in real-world scenarios, resulting in limited accuracy in practical applications. To solve these problems, this paper propose a new framework based on deep learning and iterative optimization algorithms, achieving accurate object recognition and 6-DoF pose estimation. The framework is divided into a stacked object segmentation module, an object classification module, and a pose estimation module. In the object segmentation module, we introduce a representation method based on Euclidean distance and point cloud density to address the under-segmentation issue present in the LCCP algorithm. The target classification module is improved on the basis of PointNet +  + network, using the point cloud output by the segmentation module as the input of the classification module, and outputting the target category and mask. The pose estimation module achieves accurate 6-DoF pose estimation of the target by registering the segmented candidate point cloud with the corresponding 3D model of the target category, based on the improved SAC-IA-ICP algorithm. The proposed framework model undergoes extensive testing on two datasets, with experimental results showing superior performance in 6-DoF object pose estimation.</p>