Virtual sample deep learning-based robot target identification and pose reconstruction method

Abstract

The invention provides a virtual sample deep learning-based robot target identification and pose reconstruction method. The method comprises the steps of extracting a region of an operation target in a camera image by adopting a CNN region detector, and preliminarily determining relative positions of the operation target and a robot end camera; estimating an observation angle deviation of a current view angle and an accurate pose solving optimal view angle of the computer end camera by adopting a CNN pose classifier; controlling a robot motion by adopting a multi-observation view angle correction method to enable the end camera to be transferred to the accurate pose solving optimal view angle; and by adopting virtual-real matching of contour features and pose inverse solving at the optimal view angle, realizing accurate calculation of a target pose. According to the method, the problem in massive sample demands of a deep convolutional neural network is solved, and the problems of feature shielding and matching difficulty caused by an excessively large contour matching view angle deviation are solved; and the activeness of robot vision perception and the algorithm flexibility of target pose reconstruction are improved.

Description

technical field [0001] The present invention relates to the fields of machine vision and robot control, in particular to a robot target recognition and pose reconstruction method based on deep learning of virtual samples. Background technique [0002] At present, the visual perception of industrial robots mainly uses the key contour features on certain planes of the target workpiece to reconstruct the target pose and plan the operation path. The disadvantages include: when the deviation of the viewing angle of the object is too large, there will be problems of feature occlusion and difficult matching. ; The software is not flexible enough. For different operation targets, different contour features and corresponding pose inversion formulas need to be defined in advance. [0003] After searching: Wang Zhongren et al. proposed to use CAD model for template training in "CAD model-based random workpiece visual recognition and positioning method", and then use the template outlin...

AI Technical Summary

Problems solved by technology

[0002] At present, the visual perception of industrial robots mainly uses the key contour features on certain planes of the target workpiece to reconstruct the target pose and plan the operation path. The disadvantages include: when the deviation of the viewing angle of the object is too large, there will be problems of feature occlusion and difficult matching. ; The software is not flexible enough. For different operation targets, different contour features and corresponding pose inversion formulas need to be defined in advance.

[0004] However, the above method does not have the initiative of viewing angle selection. If the current camera pose corresponds to less contour information of the target image, it is difficult to achieve accurate matching; and the traditional template matching method cannot adapt to the contour noise introduced by the complex background in the actual industrial scene. Moreover, the template search and nonlinear optimization process will take a lot of time, which cannot meet the real-time requirements of industrial control

Images