Research On Robot Intelligent Grasping And Object Pose Estimation In Accommodation Space

Robots and related technologies are among the key technologies of “Made in China 2025”.Facing to continuous industrial upgrading,industrial robots need sensors such as cameras to guide them to complete the operation of random rigid objects in the scene.Compared with the rigid objects,as the changeable shape or size of flexible object,machine vision can't accurately detect images to extract the key feature points,which will lead to pose bigger error or incorrect results.Especially in the manufacturing process of flexible semi-finished or finished products,workers need handling,loading and unloading of the inner semi-enclosed space of container online storage in production,testing,storage link.According to theory and application of machine,this paper studies pose estimation of flexible object,trajectory optimization,pose estimation in accommodation space,etc.Based on machine vision model,this paper introduces position measurement method of known object,and then proposes improved binocular calibration based on ICP algorithm,binocular sparse point cloud based on P-KLT algorithm,binocular stereo calibration based structure light and the dense point cloud.In the preprocessing method of point cloud,point cloud segmentation and initial registration method(based on camera posture and image feature points)based on image data are proposed.In trajectory planning,joint space trajectory planning and cubic b-spline curves are used to propose a trajectory planning method based on quantum genetic algorithm and improved fruit fly optimization.In this paper,pose estimation method based on improved fruit fly optimization algorithm is proposed for the placement of objects in accommodation space.The specific research contents of this paper include: 1.3D objects measurement and pose estimation based on machine visionIn this paper,the analytical geometric measurement model of monocular vision is introduced in detail.In order to obtain the measured workpiece information more accurately,it is necessary to calibrate the pose of binocular camera coordinate system accurately.The external parameters of binocular vision and the same target plane were used for obtaining binocular vision pose matrix,which would be compensated by using the rotation and translation matrix of the two groups of point set based on ICP algorithm.After binocular vision calibration,Harris operator was used to detect the segmented target area and p-klt tracking algorithm was used to track the feature points in the right image to obtain the sparse point cloud of the target object.The structural light system is another method to measure the surface information of 3d space objects.Monocular structured light system uses orthogonal gray code grating to obtain dense point cloud.The acquired point cloud generally includes the target and field.In the application process of point cloud,the point cloud segmentation algorithm is applied to obtain the target point cloud.Then,it is necessary to estimate the pose between the target point cloud and the model point cloud,or to fuse multiple point clouds.Based on corresponding relationship between image pixel coordinates of monocular vision and the position coordinates of point cloud,the same physical scale of the world coordinate system and the point cloud and rotated point cloud,this paper proposed three point cloud segmentation methods: based on the world coordinate system of the target,based on image threshold and based on perspective transformation.Finally,the three point cloud segmentation methods were tested to compare the results.Two methods of initial point cloud registration were proposed,based on camera pose and image feature points.Rotation matrix and translation matrix were obtained by using reference images taken by cameras before and after moving,SURF operator,bipolar line constraint,internal parameters and feature point depth value.Two methods respectively established depth estimation model of binocular vision and monocular vision structured light system,including the camera pose estimation methods using residual rough matching point,image feature points methods removing grass error to obtain SVD solution accuracy.2.Optimal trajectory planning and grasping of the robotIn industrial field,the execution path of the robot is usually complicated and changeable.In a fixed path or multiple paths,the goal of robot's execution is to obtain the optimal time or consume the least energy.In this paper,we first establish cubic B spline of data points,and under constraints of the velocity,acceleration and the second accelerations of the robot dynamics,proposed the trajectory optimization algorithm based on quantum genetic algorithm and fruit fly algorithm respectively.In order to apply quantum genetic algorithm to trajectory optimization,fitness function needs to be adjusted.Variable fitness function is used to obtain effective trajectory planning results after iterative convergence.The population optimization of fruit fly was conducted by individual sequencing,and half of the optimal population randomly generated the new half group,which was combined into the next iteration group.Compared with the quantum genetic algorithm,the optimized results show that the algorithm can obtain better target function values,and verify the reliability and practicability of the algorithm.3.Pose estimation and grasping of the object in accommodation spaceThe accommodation space changes as flexible products are packed into it.The goal of this study is to establish a method for pose estimation of a target object in the accommodation space.The section introduces basic algorithms and concepts,including the quick hull(Qhull)algorithm,oriented bounding box(OBB)algorithm,and fruit fly optimization algorithm(FOA).The constraint conditions and the objective function of pose estimation are set up according to the pose variables in three-dimensional(3D)space,and a solution method for pose estimation is established using an improved FOA.Based on proposed method,an industrial robot scene is constructed and the acquisition process of point cloud based on binocular vision is introduced in detail,and pose matrix of grasped object is solved.Under constraint of object gravity and grasping pose in the process of actual loading situation,the algorithm is further optimized and improved.The experimental results show that the proposed pose estimation in accommodation space could search for an optimal pose(6 degrees of freedom)and guide robot to place object,and the feasibility of the method is validated.