基于非人工标注图像的开放场景中物体6D姿态估计方法

研究准确率高、泛化能力强的物体6D姿态估计方法是计算机视觉应用的迫切需求，也是亟需解决的共性关键科学问题。项目聚焦于非人工标注图像样本条件下，面向开放环境的物体6D姿态估计问题，开展以下研究：1）发展现有基于3D模型生成合成图像的6D姿态估计方法，该方法是目前唯一的基于非人工标注图像的方法，重点解决现有方法中合成图像与真实图像之间不可避免的差异所导致的准确度下降问题；2）开辟一条基于非人工标注图像样本的物体6D姿态估计新途径，即使用真实物体拍摄照片及其拍摄视点之间的相对变化作为训练样本，通过解决拍摄视点间相对变换与物体姿态之间没有直接联系这一核心问题，实现姿态估计；3）将图卷积网络引入6D姿态估计方法中，重点解决利用物体自身几何、语义联系预测物体关键点的问题，并将其嵌入到上面两种方法中，进一步提升这两种方法的泛化能力和准确率。项目成果将实际应用于申请人团队已开发的增强现实博物馆游览系统中。

Research on the 6D pose estimation method of objects with high accuracy and strong generalization ability is an urgent need for computer vision applications, and also a common key scientific problem that needs to be solved urgently. This project focuses on the problem of 6D pose estimation of objects in an open environment under the condition of non-manually labeled image samples. The following research is carried out: 1) The development of existing 6D pose estimation methods based on 3D models to generate synthetic images. This method is currently the only The method of manually labeling images focuses on solving the problem of reduced accuracy caused by the unavoidable difference between the synthesized image and the real image in the existing methods; 2) A new approach to 6D pose estimation of objects based on non-manually labeled image samples is developed, namely Using real objects to take photos and the relative changes between the shooting viewpoints as training samples, by solving the core problem that there is no direct connection between the relative transformations between shooting viewpoints and object poses, to achieve pose estimation; 3) Introduce graph convolution network The 6D pose estimation method focuses on solving the problem of predicting key points of objects using their geometric and semantic connections, and embedding them in the above two methods to further improve the generalization ability and accuracy of the two methods. The proposed method will be applied to the augmented reality museum tour system developed by the applicant team.