基于可变形卷积的双目视觉三维重建  

Binocular Vision 3D Reconstruction Based on Deformable Convolution

作　　者：李鹤喜 李威龙 LI He-xi;LI Wei-long(Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen,Guangdong 529020,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020

出　　处：《计量学报》2022年第6期737-745,共9页Acta Metrologica Sinica

基　　金：广东省自然科学基金(2016A030313003)。

摘　　要：提出一种基于可变形卷积的立体匹配算法来进行双目视觉三维重建。首先,采用二维可变形卷积对输入的左右两幅图像进行特征提取;然后,利用三维可变形卷积,在匹配代价空间中有效地聚合两个图像之间的相关特征;最后,采用3个阶段级联残差学习的方式来降低匹配代价空间的参数计算量,以达到快速匹配的实时要求。根据该算法原理完成了视差深度图的检测,并通过Open3D重建三维物体。实验结果表明:该算法的参数量为0.5×10^(6),运行时间只需0.02 s,生成的视差图精度较高,三维重建效果较好。A stereo matching algorithm based on deformable convolution is proposed to perform 3D reconstruction of binocular vision.Firstly,the two-dimensional deformable convolution is used to extract the features of the left and right input images.Secondly,the three-dimensional deformable convolution is used to effectively aggregate the relevant features between the two images in the matching cost volume.Finally,a three-stage cascade residual learning method is used to reduce the parameter calculation amount of the matching cost volume,which can meet the real-time requirements of fast matching.According to the principle of the algorithm,the detection of the disparity depth map is completed,and the three-dimensional object is reconstructed through Open3D.The experimental results show that the parameter amount of the algorithm is 0.5×10^(6),the running time is only 0.02 s,the generated disparity map has high precision,and the reconstructed 3D effect is good.

关 键 词：计量学 双目视觉 可变形卷积 三维重建 立体匹配 

分 类 号：TB96[机械工程—光学工程]