Study On Object 6-DoF Pose Estimation Problem For Robotic Grasping Scene

In an autonomous robotic grasping scene,the robot needs to obtain the target object 6-Do F pose to determine the grasping pose and generate the motion trajectory.A fast and accurate target object 6-Do F pose estimation can efficiently improve the grasping success rate in many situations,such as object occlusion,background clutter,and illumination changes.Target object 6-Do F pose estimation refers to the transformation relationship between the target object coordinate system and the camera coordinate system.When the coordinate system origins are different,the key to improving the accuracy of the 6-Do F pose estimation is properly addressing the non-linearity of the rotation space.Meanwhile,the background objects and lighting conditions have a large difference in different robotic grasping scenes.To successfully grasp the target object,CNN is applied to extract features from the scene and train it in different environmental pose estimation datasets,which make the algorithm adapt to different grasping scenes.Therefore,this paper adopts RGB-D images with rich scene information as input,and studies the target object 6-Do F pose estimation algorithm based on CNN.The algorithm first performs edge-point detection,and then applies point set registration to obtain the object pose.By expanding the data set,the accuracy of the pose estimation of the real grasping scene is improved,and ultimately improve the success rate of the robot grasping the target object.The main works are:(1)The production of 6-DOF pose estimation dataset based on real scene is proposed to adapt the algorithm to the realistic robot grasping environment.A 6-Do F pose estimation dataset production system based on real scenarios is designed,which introduces a viewpoint point uniform generation algorithm and has the ability to use the object 3D model to automatically generate 6-Do F pose estimation dataset containing real scene lighting information and background objects.(2)A new two-step method for 6-DOF pose estimation of objects is proposed.In the training process of the first step edge-point detection,an automatic key point detection algorithm is proposed for the disorder of the point cloud,which effectively reduces the error of the edge-point detection by optimizing the algorithm training process and the time consumption of the algorithm is decreased by filtering out the background points.Then,the object pose estimation is obtained through the edge-point set registration in the second step.The algorithm finally achieves the improvement of pose estimation accuracy and time efficiency.(3)To verify the effectiveness of the algorithm,both simulated and real grasping experiments are designed for the robot.The algorithm accuracy in estimating target object 6-Do F pose is demonstrated through the simulation grasping environment.Furthermore,the real grasping experiment verifies that the proposed 6-Do F pose estimation algorithm is available for the autonomous grasping task of industrial robots.