3D Object Detection And Pose Estimation For Robotic Manipulation

3D object detection and pose estimation are of great significance to robotics research area because they enable robot to autonomously perform manipulation tasks,such as pick-n-place,parts assembly,etc.However,as robots are getting more deployed in unstructured environments,cluttered scenario still pose challenges to existing algorithms.This paper,taking RGB image and depth image as input,studies 3D object detection,pose estimation and robotic grasping in general environment.The main contents of the paper are as follows:Proposing a template clustering algorithm to address the problem of duplicate results and high false positive rate,which are common in template matching methods.Based on templates' spatial location,we cluster templates into groups.Then we design a scoring function to evaluate all the clusters and perform non-maximum suppression in image to remove duplicate results.Meanwhile,we filter part of the clusters based on their sizes.Proposing a pose estimation pipeline based on the detection result.We first recover an object hypothesis' s initial pose using templates' training pose.At pose refinement scheme,we present a coarse-to-fine ICP registration by dynamically changing ICP's distance threshold.Finally,we propose a hypothesis verification algorithm measuring the overlap degree between the model points and the scene points.Designing and implementing an automated hand-eye calibration system.Once the calibration board is placed in the robot's workspace and a desired number of poses are input,the system automatically drives robot to multiple poses,saves the robot poses and the corresponding calibration board poses,and finishes the computation.Introducing a model-free grasping pose generation scheme for suction gripper.We first use region growing algorithm to segment an object's corresponding scene points,acquiring the largest smooth surface for suction.Then we compute the centroid and its normal to generate a grasping pose.Building a robotic bin-picking system.The hardware of the system consists of a 3D camera,an RGB camera and an UR5 robot.The software of the system is built on top of Robot Operating System.Based on a sense-plane-act logic,the system implements an autonomous process of random object picking,verifying the feasibility of all the proposed methods.