面向目标6D姿态追踪的复用预测网络
Reusable predictive network for target-oriented 6D attitude tracking
呼木吉力吐1
作者信息
- 1. 国能准能集团有限责任公司 科学技术研究院,内蒙古 鄂尔多斯 010300
- 折叠
摘要
为提高目标6D姿态追踪网络的收敛能力和追踪精度,提出一种基于少量数据驱动的目标 6D姿态追踪复用预测网络.以当前时刻的彩色及深度(red green blue and depth,RGB-D)图像和上一时刻的目标渲染值作为输入,通过2 个独立的特征编码器提取特征矩阵,在特征编码器中引入通道注意力机制模块,保证有选择性地调整通道信息的权重;构建复用预测网络模块,将特征矩阵解耦得到旋转矩阵,通过旋转矩阵前向传播与特征矩阵融合,将融合的结果再次解耦得到物体6D姿态的旋转矩阵与平移矩阵,并采用李代数方法通过2 个矩阵计算出目标的 6D 姿态.实验结果表明:在使用少量数据训练网络模型的情况下,与 MaskFusion、"TEASER++"和se(3)-Tracknet等方法相比,所提方法能够提高目标6D姿态追踪的准确率.
Abstract
To improve the convergence ability and tracking accuracy of the target 6D attitude tracking network,a target 6D attitude tracking multiplexing prediction network driven by a small amount of data is proposed.Using the RGB-D image at the current time and the target rendering value from the previous time as inputs,feature matrices are extracted through two independent feature encoders.A channel attention mechanism module is introduced in the feature encoder to selectively adjust the weights of channel information.A reusable prediction network module is constructed to decouple the feature matrix into a rotation matrix,then propagate the rotation matrix and fuses it with the feature matrix through forward propagation.The fusion result is decoupled again to get the rotation matrix and the translation matrix of the object's 6D attitude,and the Lie algebra method is used to calculate the target's 6D attitude through the two matrices.Experimental results show that when training network models with a small amount of data,our method can improve the accuracy of target 6D pose tracking compared with methods such as MaskFusion,TEASER++,and se(3)-Tracknet.
关键词
6D姿态追踪/深度学习/神经网络/数据驱动/注意力机制Key words
6D pose tracking/deep-learning/neural networks/data-driven/attention mechanism引用本文复制引用
基金项目
辽宁省教育厅基本科研项目(LJKMZ20220677)
出版年
2024
国家科技期刊平台
NSTL
万方数据