改进特征提取网络的6D目标姿态估计算法

An improved 6D target pose estimation algorithm for feature extraction networks

为解决弱纹理目标和遮挡目标难以进行实时准确姿态估计的问题,在DenseFusion的框架上提出改进特征提取网络的6D目标姿态估计算法:(1)在图像特征提取阶段加入跳跃连接方法及注意力机制模块,使得深层特征与浅层特征有效融合,提高特征信息的丰富性和有效性;(2)在点云特征提取阶段使用PointNet进行点云的初始特征提取,采用K近邻法和全局池化,获得更丰富的点云特征信息;(3)将图像特征和点云特征进行密集融合进行姿态估计和优化。实验表明,改进后的算法在LineMOD数据集和Occlusion LineMOD数据集上的表现都优于DenseFusion算法,且改进的图像特征提取网络和改进的点云特征提取网络不论是单一使用还是叠加使用均能有效提高姿态估计的准确率。

In order to solve the problem of real-time and accurate pose estimation for weakly textured and occluded targets,a 6D target pose estimation algorithm based on an improved feature extraction network is proposed in the DenseFusion framework.Firstly,in the stage of image feature extraction,skip connections and attention mechanism modules are added to effectively fuse deep and shallow features,improving the richness and effectiveness of feature information;Secondly,in the point cloud feature extraction stage,PointNet is used for initial feature extraction of the point cloud,and then K-nearest neighbor method and global pooling are used to obtain richer point cloud feature information;Finally,the image features and point cloud features are densely fused for pose estimation and pose refinement.The experimental shows that our method outperforms DenseFusion on the LineMOD dataset and Occlusion LineMOD dataset,and the improved image feature extraction network and the improved point cloud feature extraction network can effectively improve the accuracy of pose estimation whether used alone or in combination.