Abstract: To solve the difficulty of obtaining accurate position and pose information of objects in complex scenes with low resolution depth camera, an object pose estimation method based on point cloud scene segmentation and improved registration algorithm is proposed. Firstly, a structured light scanning 3D scanner is proposed to make the template of the object, to eliminate the difference caused by the template directly generated by the theoretical model. Then, an object segmentation method based on two-step method is proposed, which can quickly and accurately complete the segmentation of the object in the scene cloud. Finally, a key point extraction algorithm that combines the normal angle constraint and the neighborhood number constraint is proposed, which can effectively extract the key points with large curvature characteristics and non-noise in the template and scene point clouds, and then calculate FPFH description at the key points, and complete object coarse registration and initial pose estimation based on random sampling consistency. To improve the accuracy of pose estimation, an improved ICP algorithm with normal angle constraint is adopted to complete the accurate correction of the initial pose estimation. The proposed method is verified by experiments. Compared with the existing pose estimation methods based on point cloud registration, the error of pose estimation is significantly reduced, which effectively proves the feasibility of the proposed method.

Key words: pose estimation, scene segmentation, point cloud registration, ICP algorithm, feature constraints