Depth 6D pose estimation network model and workpiece pose estimation method

Abstract

The invention discloses a depth 6D pose estimation network model and a workpiece pose estimation method, the workpiece pose estimation method solves the problem that a large-volume point cloud data set is difficult to acquire through a method of generating a simulation data set by a physical engine, and a complete scene point cloud is directly used as input by a pure geometric point cloud coordinate. Through the semantic and instance segmentation part of the point cloud, the local and global features of the input point cloud can be extracted, the ability of the network to understand the scene is improved, the accurate attitude is output through the multi-layer feature fusion pose estimation network, the problems of object stacking and self-shielding are solved to a certain extent, the robustness is provided for various symmetrical objects, and the method is suitable for large-scale popularization and application. Through experimental verification of simulating a data set of a real scene, the method provided by the invention has obvious advantages in overall precision and stability, and has higher robustness.

Description

technical field [0001] The invention relates to the technical field of machine vision, and specifically provides a deep 6D pose estimation network model and a workpiece pose estimation method. Background technique [0002] With the advancement of technology, robots are increasingly widely used in grasping, assembly, packaging, processing, logistics sorting, etc. Among them, the grasping and assembly operation is the most common application scenario of robots, and a major challenge is the accurate grasping of objects in complex environments. In the traditional structured environment, by manually searching for the teaching point and programming it into the program, and then running according to the fixed program, although high accuracy and success rate can be achieved, due to the lack of environmental awareness and interaction, it is difficult Complete precision grasping and assembly operations in complex scenarios such as structured or semi-structured. [0003] A reliable r...

AI Technical Summary

Problems solved by technology

[0004] Most industrial parts are rigid objects with standard production models, and there is almost no change within the class, which reduces the difficulty of identification. However, most parts in industrial scenes have weak textures, similar or identical colors. In this case, RGB color information is no longer reliable

Therefore, in order to improve the reliability of recognition, this paper starts with the point cloud that only contains 3D geometric information, and mines the 6D pose of the object from the edge information and geometric relationship of the object; on the other hand, the object to be captured is often in a messy scene. , mutual occlusion and stacking between objects are inevitable, which still brings great difficulties to tasks such as vision-guided robot grasping

Images