Convolution neural network-based approach to target pose recognition for robot control

Abstract

The invention discloses a robot control target position and posture identification method based on a convolution neural network, which comprises the following steps: (1) collecting image data of different positions and postures of the control target by a binocular camera to form a sample data set; (2) labeling the sample data set; (3) Constructing deep convolution neural network model; (4) collecting a new image sample, and obtaining the pixel coordinates of the feature points of the new image sample by using the deep-layer convolution neural network model; (5) obtaining a projection matrix corresponding to the binocular camera; 6) obtaining that three-dimensional coordinates of the feature point correspond to the pixel coordinates of the feature points; (7) mapping The three-dimensional coordinate transformation of the feature points to the robot control coordinate system to obtain the position and posture information of the control object. The invention not only fully utilizes the object characteristic information, but also fully considers the influence of external disturbance, and simultaneously avoids the problem that the position and attitude information are difficult to be calibrated by the common depth neural network, and improves the robot control target attitude detection technology.

Description

technical field [0001] The invention belongs to the technical field of robot control, and in particular relates to a method for recognizing the pose of a robot control target based on a convolutional neural network. Background technique [0002] In the process of modern industrial production, with the widespread application of industrial robots in the industrial field, some dangerous working environments or simple and repetitive mass industrial production operations that are not suitable for manual work have been gradually replaced by machine operations. In industrial manufacturing sites, industrial robots can achieve efficient grasping and sorting operations on products, but the accuracy of machine operations is limited by the detection accuracy of product recognition and positioning by machine vision systems, attitude estimation, etc., and it is difficult to meet industrial production applications. , thus limiting the popularization and application of industrial robots, it...

AI Technical Summary

Problems solved by technology

Among them, the algorithm based on the extraction of color features is simple and invariant to scaling, rotation and translation, but the lack of color information lies in ignoring the spatial position relationship in the image, and the color features are greatly disturbed by external lighting factors.

The object detection method based on texture features can obtain object structure and spatial information very well, but for some industrial products, the surface texture features are not obvious, and the texture feature-based detection method is not very applicable

The classification method based on the convolutional neural network is self-adaptive and can self-learn to extract the features required for classification. However, this method requires a large amount of labeled data, and there is a problem that it is difficult to accurately calibrate information such as position and attitude manually.

Images