Annotation

Annotation: The article presents a method for segmenting a point cloud of a scene consisting of unknown objects based on the use of the Voxel Cloud Connectivity Segmentation (VCCS) method and two-stage feature vector processing using the PointNet neural network and Dynamic Graph Convolutional Neural Network (DGCNN). In some cases, the practical application of manipulative robots involves grasping objects whose shape, color and other features are not known in advance. In particular, examples of such tasks can be cleaning of premises, emergency rescue operations to remove blokage, work in warehouses or in post offices. In the proposed approach, an image of the scene in the form of a point cloud is compiled from a set of images of a cluttered scene obtained from RGBD cameras, then this point cloud is processed using the VCCS heuristic algorithm and machine learning methods. The result of the approach is a segmented point cloud, for each point of which there is a label that determines its belonging to a separate object in the scene. The novelty of the approach lies in the combination of the VCCS heuristic algorithm and the new neural network architecture, which is a combination of modified PointNet and DGCNN networks. The conducted experimental studies confirm the operability of the proposed solution.