To enhance the efficiency and accuracy of partial-to-whole point cloud registration for complex workpieces, this paper presents a novel method based on the Mean Feature Descriptor (MFD). The proposed approach extracts geometric features from key points in the scanned point cloud, constructs local feature descriptors using a localized coordinate system, and performs coarse registration. Compared to conventional local descriptor-based methods, the MFD algorithm not only effectively captures local geometric characteristics but also significantly improves computational efficiency while maintaining high registration accuracy. Experimental results demonstrate that the MFD-based method substantially accelerates registration and measurement processes for complex workpieces. It exhibits strong robustness against noise and varying point cloud resolutions, outperforming existing descriptors such as PFH, FPFH, and HoPPF in terms of F1 score and matching precision. The method achieves reliable registration even under challenging conditions, such as partial overlap and geometric feature sparsity. Notably, the MFD descriptor inherently captures geometric symmetry-invariant features of local point cloud regions especially symmetric interfaces of complex workpieces. This ensures stable registration performance even when partial scans only preserve part of the workpiece’s symmetric structure.

Accurate modeling of multi-part objects with diverse quasi-conical and cylindrical geometries requires a unified understanding of both their semantic composition and underlying geometric structure. To address this challenge, a joint Pose and Parametric Geometry Estimation (PPGE) framework is proposed, which integrates learned representations with explicit geometric priors for simultaneous segment-wise segmentation and parametric geometry prediction from raw point clouds. The proposed hybrid architecture employs an augmented PointNet++ backbone to extract multi-scale spatial features and a squeeze-and-channel-attention module to enhance inter-channel discriminability and context awareness. Three slot-based task-specific prediction heads are designed to separately infer segment membership, symmetry axis orientation, and point-wise surface normals, providing complementary semantic and geometric cues. These multi-branch outputs are further coupled by a geometrically-constrained probability-axis-normal fusion module, which enforces alignment between semantic segmentation, axis direction, and normal consistency, yielding coherent segment reconstructions even under partial or noisy observations. To refine parametric accuracy, a mode-climb parameter estimation module iteratively adjusts axis direction, height and radius parameters toward physically valid local optima. Experimental evaluations on multiple datasets of quasi-conical and cylindrical assemblies, along with physical prototype validation, demonstrate that PPGE achieves superior robustness and accuracy compared to baseline and ablated variants, particularly under challenging conditions of segment overlap, noise, and occlusion.