This paper addresses the practical needs and challenges of workpiece point cloud registration in industrial environments and proposes an efficient, non-learning-based point cloud registration framework. First, to address point cloud noise and outliers commonly found in industrial environments, an adaptive outlier rejection mechanism is designed to enhance the algorithm’s adaptability to complex noise. Second, to address the complex surface structures and large initial pose errors of workpieces, a multi-scale feature fusion registration method is designed. By extracting multi-level geometric features such as edges, corners, and planes, the robustness of the registration is effectively improved. Finally, to address the physical constraints inherent in industrial workpiece registration, an improved ICP algorithm based on physical constraints is proposed. This algorithm incorporates physical information such as the workpiece’s dimensional tolerance and assembly orientation into the point pair matching and optimization process, improving the physical feasibility and accuracy of the registration. Experimental results demonstrate that the proposed method achieves excellent performance in multiple industrial workpiece point cloud registration tasks, outperforming traditional methods in both accuracy and robustness. This method provides strong technical support for point cloud registration in industrial automation and intelligent manufacturing.