High Frame Rate Binocular Vision Based Object Pose Estimation Method And Its Application In Robot Grasp

Three-dimensional object detection and pose estimation allow the robot to perceive the presence of a target object and calculate its position and posture in an unstructured three-dimensional space.It is the basis for the robot to operate any object,and it has significant importance for both industrial robots and service robots.The rapid development of the robot makes its working environment gradually move from a single rule scene in traditional industrial production to a more dynamic unstructured scene.This undoubtedly greatly increases the three-dimensional object detection,pose estimation speed and accuracy requirements.At present,there are two ways to estimate poses of three-dimensional objects.One is based on two-dimensional image retrieval,and the other is based on three-dimensional point cloud registration.The former needs to obtain a large number of two-dimensional images of objects at different viewing angles,and the memory consumption is huge and it is difficult to achieve high accuracy.The latter has better robustness and flexibility,and can achieve higher accuracy,and has become a trend for applications requiring high-precision pose estimation.However,its performance is heavily dependent on the accuracy of the depth sensor,and pose estimation from a single depth map may fail when no three-dimensional features are encountered.Most current studies have chosen to use depth cameras directly,such as depth cameras based on structured light and time of flight(ToF),but these active light depth cameras are severely affected by ambient light and may interfere with each other when multiple devices are used simultaneously.In contrast,the binocular camera does not have the above problem,but the algorithm of binocular vision is much complicated and it is difficult to achieve real-time performance,not to mention the accuracy is relatively low.In this paper,the following work has been done from aspect of depth map acquisition and the use of multi-view fusion when estimating the pose:(1)A high frame rate binocular matching algorithm framework based on multi-scale weighted voting algorithm is proposed.The algorithm framework is suitable for a variety of stereo matching algorithms.This paper uses the simplest block matching algorithm as an example to verify the effectiveness of the algorithm framework.(2)The TSDF Volume model is used for timing fusion of continuous depth maps,which improves the accuracy of the depth map measurement and improves the integrity of the depth map.(3)The coarse-to-fine point cloud registration method that combines FPFH three-dimensional feature and ICP was used to achieve registration of two-frame point clouds with large displacement,Finally realizing the highprecision pose estimation of three-dimensional objects in the robot working space under different posture states.(4)A robot grasping system was built.In the experiment,postures were estimated and captured for different poses of various objects.In the experiment,the pose estimation and grasping are carried out for various objects in different positions.The success rate of object detection is 100%,the accuracy of the position estimation is 0.5 mm,the direction accuracy is 1 degrees,and the success rate of grasping is 90%.The accuracy and precision of the proposed algorithm are fully verified.