Experimental Verification of Machine Vision-Based Pose Recognition Method for Bell-and-Spigot Joint Shield Segments

Target pose recognition is a critical component of automatic assembly for bell-and-spigot joint shield segments, providing essential visual information for the precise control of segment assembly robots. This paper introduces a robust target pose recognition method, focusing on two core phases: segment grasping and positioning. During the grasping phase, point cloud registration is performed using a multi-level kernel point convolution and attention mechanism. Preprocessing steps, including octree-based clustering and noise removal, ensure efficient key point extraction while preserving geometric integrity. Hierarchical feature extraction combines local and global geometric relationships, enhanced by a multi-level self-attention module, enabling precise pose estimation for segment grasping. In the positioning phase, a 3D U-Net model extracts critical feature clusters, such as inner arc surfaces and grooves, from the segment point cloud. Coarse-to-fine matching, combined with constraints on all six degrees of freedom, produces an accurate pose transformation matrix, ensuring precise alignment with the segments in the previous ring. A full-scale test platform for automatic segment assembly was constructed, and experimental results demonstrate the robustness and reliability of the proposed method, achieving high-precision pose recognition and seamless segment assembly.