Research On Detection Of Grasp Poses Based On Point Cloud Camera

With the improvement of the technology level of visual sensing equipment and the deepening of the image processing algorithm research,the level of automation and intelligence of robot is increasing.The visual based grasping system is used widely in the field of industrial robots for its several advantages such as high efficiency,flexibility,etc.In the process of grasping,the system is underperforming for the diversity of positions and poses of the object.In addition,the shooting angle of the point cloud camera,the shape,type and size of the object to be grasped are all uncertain.Therefore,it has important theoretical significance and practical application value to study the generation method of grasping pose.In the view of the grasping pose uncertainty of the object,the main work completed is as follows:Consider the shape,size,type and placement uncertainty of the object,in the situation that the point cloud camera shots from top to bottom,the depth image of the object to be grasped is obtained by using the point cloud camera.The gradient boundary of the object to be grasped is extracted by Sobel operator and the gradient boundary is obtained by using Alpha-Shape algorithm.Then,K-Means algorithm is adopted to cluster the depth of the object to be grasped into three depth values,which are used to replace the original depth data.Grasping candidates satisfying the condition are sampled based on the maximum opening width of the parallel jaw and the force closure restriction.By adopting Grasp Quality Convolutional Neural Networks(GQCNN)training by Dex-Net2.0 data set,grasp candidates are graded and ranked.On the basis of the previous section,considering the defect of the algorithm when the object to be grabbed deviates from the center of the desktop,a multi-directional grasp pose generation algorithm is designed.In this algorithm,the point cloud camera is placed at a fixed tilt angle to obtain information about the side of the object to be grasped.Then point clouds which is independent with the object to be grasped is removed by using the random sampling consistency method and the noise caused by reflection and other reasons is removed by using statistical filtering method.Center points of grasp candidates are randomly selected from the surface of the point cloud image.The orientations of grasp candidates are selected from a discrete set of orientations(6 in our implementation)according to several geometric conditions.Lastly,depth images according to grasp candidates which are obtained by point cloud projection are sent into the GQCNN to be graded and ranked.Aubo manipulator and Robotics two-finger fixture is used to carry out the experimental research for verifying the two designed grasping pose generation algorithms.The experiment results show that the designed two algorithms can accomplish the grasping task effectively.