6-DOF grasping pose detection method incorporating instance segmentation

To improve the accuracy and reliability of the grasping pose generated by the grasping detection model, in this paper, the GPD(Grasp pose detection in point clouds) is targeted at the problem that rough object detection may treat multiple objects as one object to generate wrong grasping pose and fail to generate grasping pose at the edge of point clouds with thin walls. A grasping pose detection method combining instance segmentation and sampling based on point cloud normal vector and principal curvature was proposed. First, the Yolov7 instance segmentation network is used to generate object masks for RGB images collected by the KinectV2 camera and crop out the point cloud where the object needs to be grasped. The sampling candidate grasping the pose of the point cloud is performed based on the normal vector and principal curvature geometric features of the point cloud. Furthermore, PointNet++ was used to evaluate the network classification of grasping pose by feature extraction. Finally, collision detection was carried out on positive tag grasping pose to screen out non-collision grasping pose. Experimental results show that the grasping pose detection method integrated with instance segmentation in this paper has a higher success rate than GPD in generating a feasible grasping pose, and can solve the problem that GPD cannot generate the grasping pose at the edge of point cloud with thin walls.