Object Recognition And Pose Estimation For Robot Grasping

With the development of RGB-D sensors and robotics,service robots based on stereo vision are more and more widely used.Because of the uncertainties such as light,occlusion and clutter in the real environment of the service robot,there is still a great challenge for the robot visual perception.Different from the traditional computer vision,the robot vision needs to estimate 6-DOF pose while recognizing the object,which provides a priori information for subsequent robot grasping or other operations.In addition,in order to interact with human better,the realtime performance of system is also one of the key factors to be considered.In terms for the above problems,researchers proposed multi-view algorithms based on single-view recognition using the mobility of robots.However,the surface information of objects can not still be obtained from different perspectives for complex stacking scenes and the robot motion or the choice of view point is often constrained in real applications.For the task of robot grasping,this thesis studies the problems of recognition and localization in cluttered and occluded scenes based on the 3D point cloud.Considering the accuracy and real-time of recognition process,we propose a coarse to fine recognition framework which can recognize objects quickly and ensure the accuracy of pose estimation simultaneously.We conduct scene analysis on the basis of the result of segmentation and object recognition.Through the occlusion detection and hypothesis confidence analysis,the parsing pipeline can provide high-level scene information for robot operation.The main work of this paper is as follows:1.Propose a simple and quick approach to building model base.A generalized eigenvector is defined for recognition based on 2D image texture and 3D point cloud geometry.The object data is captured from different views with key point extraction,and the sparse feature models are built.The 3D dense model of each object are constructed based on structure from motion and the visual completeness are guaranteed through the optimization algorithms.2.Propose a coarse to fine object recognition framework based on model base and generalized eigenvector.On the basis of super-voxel clustering,the surface patches are merged according to the criteria of concavity,convexity and smoothness.The point cloud segmentation is completed with the idea of graph cut.The naive Bayesian nearest neighbor algorithm is improved with the rewritten generalized eigenvector and reduce the complexity of the recognition algorithm effectively.While coarse recognition finds K candidats quickly,fine pipeline ensures the accuracy of object recognition and 6 DOF pose estimation by multiple means such as matching points detection,geometric consistency test and model to scene verification.3.Propose a high-level scene analysis and active recognition method.Based on the range image and 3D point cloud,the detection and classification criterion of object boundary are defined and the occlusion topological relations between objects are established.According to the scene recognition process and objects occlusion relationship,each object hypothesis confidence is calculated from the respect of robot grasping.Combining with scene analysis information,the point cloud's bounding boxes with low confidence are estimated.The scene is rearranged and the objects are re-recognized with the manipulability of the robot to environment.4.Establish the software and hardware environment of robot recognition system and the simulations and experiments are conducted using the relevant algorithms.The realscenes are reconstructed according to the recognition result and 3D dense model based on Robot Operate System and Gazebo toolkit.The proposed methods are tested on the authoritative dataset and real-word environment and the robot can complete the cluttered and occluded task efficiently.In summary,this thesis presents a method of active recognition combining with visual perception,robot operation and environmental information.With no human intervention,the system may effectively deal with complex scenes and provide more accurate information for robot grabbing.The intelligent robot grasping system not only possesses the basic perceptibility,but also has some high-level environmental cognition ability,which can improve the flexibility and stability of the system.