Single-image robot disordered target grabbing method based on pose estimation and correction

Abstract

The invention particularly discloses a single-image robot disordered target grabbing method based on pose estimation and correction. The method comprises the steps: S1, generating an image data set ofa to-be-grabbed object model; S2, constructing a convolutional neural network model according to the image data set in the step S1; S3, importing the two-dimensional image of the to-be-grabbed objectinto the trained convolutional neural network model to extract a corresponding confidence map and a vector field; S4, obtaining a predicted translation amount and a predicted rotation amount of the to-be-grabbed object; S5, finding the optimal grabbing point of the object to be grabbed and calculating the measurement translation amount of the depth camera; S6, performing grabbing safety distancecorrection according to the predicted translation amount of the object to be grabbed and the measured translation amount of the depth camera, executing correction data grabbing if the correction succeeds, and entering S7 if the correction fails; and S7, repeating the steps S3-S6. The disordered target grabbing method provided by the invention has the characteristics of high reliability, strong robustness and good real-time performance, can meet the existing industrial production requirements, and has a relatively high application value.

Description

technical field [0001] The invention relates to the technical field of intelligent robots, in particular to a single-image robot disordered target grasping method based on pose estimation and correction. Background technique [0002] 6D pose estimation is a key technology required for artificial intelligence applications such as augmented reality, autonomous driving, and robot manipulation. It can help the robot grasp the target position and target direction to grasp the target. For example, in the Amazon Picking Challenge, the task of a robot picking a target item from a warehouse shelf is inseparable from fast and reliable pose estimation. [0003] Based on existing research, methods for 6D pose estimation can be broadly classified into template-based methods and feature-based methods. Traditional template-based methods first construct a rigid template of the object; then use the template to scan different positions in the input image, and calculate a similarity score po...

AI Technical Summary

Problems solved by technology

Feature-based methods can deal with occlusion, truncation, etc., but manual features require the target to have rich textures and are not robust to lighting and scene clutter, and are less robust

[0004] Although several new technologies have recently used depth information for object pose estimation and achieved good results, there are two problems: First, training deep convolutional neural networks usually requires a large amount of labeled data, including using Precise 6DOF poses for annotation of target objects

Compared with 2D detection, 3D detection based on convolutional neural network prohibits manual labeling of data, because the accuracy of manual labeling data cannot be guaranteed

Therefore, synthetic data can be used to train deep convolutional neural networks. Although synthetic data guarantees the accuracy of the data, one of the main disadvantages of synthetic data is the reality gap.

Second, because RGB-D cameras have limitations in frame rate, field of view, resolution, and depth range, which make small, thin, or fast-moving objects difficult to detect, active sensors on mobile devices consumes too much power

Although traditional keypoint-based methods can obtain accurate pose estimation, their applicability to robotic tasks relies on controlled environments and rigid objects with detailed information; on the other hand, CNN-based The object recognition in can get better results, such as rough pose estimation based on categories, but it requires a large amount of fully labeled training image data sets, so it will be difficult to use the CNN method to estimate the pose of actual objects

Images