Workpiece 6D pose estimation method based on deep learning

Abstract

The invention discloses a workpiece 6D pose estimation method based on deep learning. The invention relates to the technical field of robot environment perception. The method specifically comprises the following steps: collecting different workpiece images under different backgrounds and illumination conditions; constructing a semantic segmentation model to segment a target object; three-dimensional point cloud coordinates are converted into pixel coordinates for representation through a space conversion network, three-dimensional point cloud data and RGB information are fused, a dense fusionnetwork is constructed to estimate 3D position information and 3D direction information of an object, an ICP algorithm is adopted to iteratively match and finely adjust the pose, and therefore accurate 6D pose information of the object is obtained. Compared with the traditional scheme, the method has the advantages that the end-to-end 6D pose of the target object can be quickly and accurately estimated in real time in complex environments such as occlusion and disorder, and the problems of poor adaptability, low accuracy and limited real-time performance of the traditional pose estimation method in the real complex environment are effectively solved.

Description

technical field [0001] The invention relates to the technical field of robot environment perception, in particular to a method for estimating the 6D pose of a workpiece based on deep learning. Background technique [0002] The 6D pose (ie, 3D position and 3D orientation) of an object plays a key role in applications such as industrial robots, virtual reality, and automatic navigation systems. For the field of robotics, accurately estimating the 6D pose of an object is the premise and basis for precise robot grasping, which can provide position information and attitude information of objects for tasks such as robot grasping operations and motion planning. [0003] At present, when the robot performs grasping operations, the 6D pose estimation methods for objects mainly include methods based on corresponding points, methods based on templates, and methods based on voting. The method based on corresponding points is mainly to find the feature point correspondence between the i...

AI Technical Summary

Problems solved by technology

The method based on corresponding points is mainly to find the feature point correspondence between the input data and the 3D point cloud of the known object. Usually, features such as SIFT and SURF are extracted from the RGB-D data for feature matching. This method is fast, but the accuracy Difficult to guarantee

Due to complex environments such as occlusion, sensor noise, and lighting changes between workpieces, relying on manual feature extraction and fixed feature matching processes severely limits the accuracy of pose estimation, making it difficult to quickly and accurately estimate the pose of workpieces in complex environments.

The template-based method is mainly to select a similar template from the marked 6D pose template, and use its 6D pose as the current pose, but this method is mainly for images with no texture or weak texture, and needs to provide objects Accurate CAD model and size, due to the different shapes and textures of the workpiece in the real environment, it is necessary to deal with different and never-before-seen object instances in a given category, which is difficult to meet the 6D pose estimation of objects in the real environment

In the voting-based method, each pixel or 3D point obtains a 6D pose through voting. This method handles complex situations in an occluded environment better, but the real-time performance is relatively limited.

Images