一种基于生成对抗网络的相机位姿估计方法  

A Camera Pose Estimation Method Based on Generating Countermeasure Network

作　　者：陈欣然 邵帅[1] 石俊霞[1] Chen Xinran;Shao Shuai;Shi Junxia(Changchun Institute of Optics,Precision Mechanics and Physics,Chinese Academy of Sciences,Changchun,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春

出　　处：《科学技术创新》2025年第10期108-112,共5页Scientific and Technological Innovation

摘　　要：图像的拍摄角度估计对于提升机器人行为决策和物体识别的准确性具有非常重要的作用。传统的相机位姿估计方法通常使用标注数据对摄像角度进行回归学习,然而由于输入的图像经常存在几何畸变,因此会导致评估精度下降。本文提出了一种基于生成对抗网络的相机位姿估计方法,该方法将图像的潜在特征分为随视角发生变化和不发生变化的两个部分,然后将这两个部分分别输入到生成对抗网络解码器进行训练,从而消除几何畸变对姿态评估的影响。同时还利用旋转、透视后的图像对数据集进行扩充,增强网络识别的鲁棒性。实验结果表明,在标注数据较少的情况下,与现有方法相比,该方法能够显著提高相机位姿估计的精度。The estimation of image shooting angle plays a very important role in improving the accuracy of robot behavior decision and object recognition.Traditional camera pose estimation methods usually use labeled data to regress the camera angle.However,due to the geometric distortion of the input image,the evaluation accuracy will be reduced.In this paper,a camera pose estimation method based on the generated countermeasure network is proposed.The potential features of the image are divided into two parts,which change with the angle of view and do not change,and then the two parts are input to the generated countermeasure network decoder for training,so as to eliminate the influence of geometric distortion on pose evaluation.At the same time,the rotated and perspective images are used to expand the data set to enhance the robustness of network recognition.Experimental results show that this method can significantly improve the accuracy of camera pose estimation compared with existing methods in the case of less labeled data.

关 键 词：位姿估计 生成对抗网络 人工智能 机器视觉 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]