基于李代数表征的三维物体空间姿态检测

Three-dimensional object space pose detection based on Lie algebras representation

利用卷积神经网络学习并预测二维图像中三维物体的姿态信息,提出一种基于李代数的三维物体姿态表征方式。为了仅利用二维图像来准确预测三维姿态信息,采用李群和李代数将三维物体姿态分解为平移和旋转向量,姿态向量表征方式满足神经网络反向传播时要求的可微分条件,提高了训练效率。首先,通过RGBD相机获取三维物体的真实坐标信息,然后利用旋转矩阵和平移矩阵来描述物体的三维坐标,运用李代数将旋转矩阵和平移矩阵转化为对应向量,使用卷积神经网络回归对应的坐标向量来预测三维姿态信息。相比其他同类算法,本方法提升了三维物体姿态预测的准确性,提高了算法的测试速度。

In this paper,the convolutional neural network(CNN)was used in learning and predicting the pose information of the 3D objects in two-dimensional images,and a three-dimensional object pose representation method based on Lie algebras was proposed and In order to accurately predict the 3D pose information by using only 2D images,the Lie group and Lie algebras were used to decompose the 3D object pose into translation and rotation vectors.Because this pose representation method satisfied the differentiable conditions required by the back propagation of the neural network,the training effectiveness was greatly improved.Firstly,the real coordinate information of the 3D object was obtained by the RGBD camera.Then,the 3D coordinates of the object were described by using the rotation matrix and the translation matrix which were transformed into their corresponding vectors by using Lie algebras.Finally,the corresponding coordinate vectors were regressed by using CNN to predict the 3D pose information.Compared with other similar algorithms,this method can improve the accuracy of 3D object pose prediction and the test speed of the algorithm.