Résumé: The manipulation of deformable linear objects (DLOs) is an important task in many fields, which raises demands for the perception of DLOs in real situation. In this paper, we propose a cylindrical fitting reconstruction method for DLOs with only one frame of point clouds captured by a depth camera. The point clouds are first processed by operation space filtering and outlier removal to eliminate the interference. To accurately segment the specific object from the complex background, the PointSIFT module is inserted into PointNet++ architecture and fine-tuned on our dataset. To reconstruct the flexible DLOs in 3D space, an improved adaptive K-means algorithm which accommodates to the unknown length and curvature is designed. The adaptive K-means algorithm distributes the point clouds into appropriate number of cylindrical clusters. To achieve the main axis of the cylinders, we construct the point clouds covariance matrix. By applying principal component analysis (PCA), three orthogonal dimensions and the PCA bounding box are obtained. Afterward, an octree-based directional constraints is designed to sort the center points of DLOs with arbitrary curvature. The proposed framework achieves an average error of less than 1 mm during a manipulation experiment in a simulation live-line maintenance site.

Langue: Anglais
Thème Informatique

Mots clés:
3D reconstruction
Semantic segmentation
Point cloud
Deformable linear objects