Study On Texture-less Object 6-DoF Pose Estimation For Robotic Grasping

With the development of industrial technology,industrial upgrading has much higher requiements for the intelligent level of robot.The visual perception technology of robots plays a vital role as one of core modules of the intelligent robot.Object 6-DoF pose estimation is a key issue in machine vision technology.The goal is to provide the robot with the information on the operation of the target object.The 6-DoF pose is the transformation from object coordinate system to vision sensor(camera)coordinate system,which consists of 3-DoF translation transformation and 3-DoF rotation transformation.And 6-DoF pose estimation technology which can deal with complex environment is the basis of intelligent robots' autonomous grasping.In addition,some local feature descriptors have been proposed since 2000,which makes the pose estimation of objects with abundant texture information become much more easiler.However,there are many objects have no or less texture information in actual scenes,that is texture-less object.As for those texture-less object,traditional methods which based on local feature descriptors are not applicable.Therefore,texture-less object 6-DoF pose estimation in complex environment has become a research focus in recent years.This paper mainly studies texture-less object 6-DoF pose estimation methods for robotic grasping tasks,including four parts as follows:1)In order to meet the feature training requirement of the object 6-DoF pose estimation methods,a model rendering-based training data automatic generation system is realized.The system can directly import the 3D model of target object,and generate datasets including RGB-D images,RGB images and edge images.This paper utilizes electrostatic model to generate uniformly distributed sampling viewpoints.These datasets can be employed on feature training for pose estimation methods.2)As for Lambertian texture-less objects,consumer RGB-D sensors can acquire their full depth data.This paper realizes two 6-DoF pose estimation methods which are based on RGB-D images,including template matching-based holistic patch method and combined holistic and local patch method,and these methods have been evaluated on the public dataset.The first method utilizes scale-invariant RGB-D patch,hash voting-based hypothesis generation and particle swarm optimization,and the second method combines holistic patch extraction and local patch regression with convolutional auto-encoder.Experimental results show that the proposed methods have higher precision than other existing methods and can be applied on complex environment such as background clutters,foreground occlusions and multi-instance objects.3)As for non-Lambertian texture-less objects,consumer RGB-D sensors cannot acquire their full depth data.This paper realizes an edge-based 6-DoF pose estimation method which are based on RGB images,which utilizes highlight removal,edge image-based object detection and 3D edge tensor generation.The proposed method has been evaluated on the public dataset.Experimental results show that the proposed method has higher precision than other existing methods without using any depth data,and it can also be applied on complex environment such as background clutters,foreground occlusions and multi-instance objects.4)This paper builds a RGB-D sensor-based robotic grasping system and a RGB sensor-based robotic grasping system respectively.The RGB-D image-based pose estimation methods have been evalutated on the RGB-D sensor-based robotic grasping system,and the RGB image-based pose estimation method have been evaluated on the RGB sensor-based robotic grasping system.Experimental results verify that the proposed method has enough effectiveness and practicability for robotic grasping tasks.From results of public dataset evaluation and robotic grasping evaluation,the proposed methods are not only superior to other existing methods for texture-less object 6-DoF pose estimation,but also has enough practical value for robotic grasping.