Point Cloud Model Reconstruction of Deformable Linear Objects Based on Center Line Fitting

In various manipulator working scenarios, the perception of deformable linear objects (DLOs) plays a crucial role, which provides prior knowledge for grasping and obstacle avoidance. In this paper, we present a reconstruction method for DLOs based on center line fitting using a single frame of point clouds captured by a depth camera. The point cloud data undergoes preprocessing, including operating space filtering, PointNet +  + segmentation, and noise removal. To segment the data into cylindrical segments for fitting cylinder centers, an improved K-means algorithm based on B-spline interpolation is proposed, which automatically determines the number of clusters and enables rapid and uniform clustering through initial point selection using farthest point sampling. The surface normal of each segment are calculated using principal component analysis (PCA) to estimate the centers of regular cylinders. B-spline interpolation is employed again to estimate the true centerline. The final DLOs point cloud model is constructed by rotating and translating the circle point cloud model unit and superimposing it. Experimental results demonstrate that the proposed method achieves fast and accurate model reconstruction (an average error less than 2mm).