Object pose estimation method based on self-supervised learning and template matching

Abstract

The invention relates to an object pose estimation method based on self-supervised learning and template matching, and the method comprises the steps: S1, collecting a color image and a depth image ofa target object, and obtaining a corresponding color candidate image and a corresponding depth candidate image through cutting; s2, segmenting the color candidate image and the depth candidate imageby a trained self-supervised codec with a noise generator to obtain a color segmentation image and a depth segmentation image; s3, matching the color segmentation map and the depth segmentation map with a template library to obtain a matched pose; s4, finely trimming the matching pose to obtain a target object pose, wherein a color sample image and a depth sample image of the target object three-dimensional model are collected through a spherical multi-scale method; training a self-supervised codec by using the color sample graph; extracting features with pose information of the color sample map and the depth sample map under multiple scales, and constructing a template library according to the features. Compared with the prior art, the method has the advantages of being good in robustness, low in cost, free of label information and the like.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an object pose estimation method based on self-supervised learning and template matching. Background technique [0002] Object pose estimation technology is based on 3D vision to determine the 3D translation and 3D rotation transformation parameters of the target object relative to the camera, and then estimate the object pose. Object pose estimation is a key issue in the field of robot environment perception, grasping and dexterous operation. The research on this technology is of great significance to promote the promotion of service robots, industrial robot automation, VR and AR technologies. [0003] At present, object pose estimation methods are mainly based on laser point cloud, template matching and deep learning. The above technologies all have certain deficiencies, specifically: [0004] Laser point cloud-based method: collect high-precision point cloud data with high-prec...

AI Technical Summary

Problems solved by technology

This method requires a large number of samples marked with the true value of the six-degree-of-freedom pose. The success of the deep learning method depends to a considerable extent on the number of samples and the scope covered by the samples. The success of two-dimensional deep learning target detection lies in the Internet, big data A large number of samples in other fields make two-dimensional labeling easier, but samples with six-degree-of-freedom poses are quite scarce, and sample labeling is also quite difficult, so the application and promotion of methods based on deep learning is difficult

Images