Research On 3D Motion Estimation Based On Dual Quaternion

With the rapid development of digital image processing and computer vision technology, how to analyze or estimate the motion of the moving object in the three-dimensional scene from images or video sequences has gain a growing importance. Among which, since binocular vision has the capability of solving high dimension, nonlinear and non-uniqueness problem. It is undoubtedly a hot spot in the field of computer vision.Depending on stereo vision 3D reconstruction, this paper proposed a dual-quaternion-based 3D motion estimation algorithm. A large number of experiments are performed to further explore the influence pattern of feature matching on motion estimation, from which the optimization model is established, so as to the fulfillment of 3D motion estimation with better performance on accuracy and efficiency.The main contents are as followed. Firstly, 3D reconstruction based on binocular vision is studied, including camera calibration, image matching and space point coordinate calculation. Secondly, feature matching based algorithms which are widely utilized in 3D motion estimation are applied; including point/line based singular value decomposition(SVD/LSVD), orthogonal decomposition(OD/LOD) and unit quaternion decomposition(UQD/LUQD). Subsequently, to enhance the performance of the algorithms, we put forward dual quaternion for 3D rigid motion estimation in this paper, making full use of its concise mathematical expression and high efficiency advantages. A step further, we mining the key influential factor of the computation process, also through regularity summarizing and model building, an optimization model based 3D motion estimation is presented in the interest of accuracy and adaptability improvement. Finally, we testify the dual quaternion and optimization model based algorithm through whole process of experiments. The results show that our dual quaternion method can improve the efficiency of 60% with no accuracy loss. On this basis, our optimization model can further improve the accuracy of 17% with only 1% efficiency loss. Consequently, the comprehensive excellences of the proposed algorithms are proved.