基于图像的三维刚体运动估计算法比较

基于图像特征点对应估计三维刚体运动是计算机视觉的一个基本问题。对四种常用的运动估计算法,包括基于三维特征点的奇异值分解法、正交分解法、单位四元数法和基于二维特征点的"8点算法",通过大量仿真实验和一个机载目标运动光测系统进行了精度,稳定性和效率的比较。结论表明,基于三维特征的算法比基于二维特征的算法精度高、稳定性好,效率高。其中在基于三维特征的算法中,奇异值分解法效率更高,单位四元数法稳定性更好,在基于二维特征的算法中采用非线性优化是必须的,而增加特征点数和降低噪声水平对四种算法都能提高精度。讨论了这些结论对解决实际问题的作用。

A comparison of algorithms for image-based 3-D motion estimation

Motion from point correspondence is a basic problem in computer vision. Through great deal of simulation experiments and an air-loaded target motion measurement system, a complete comparison is presented among 4 kinds of motion estimation algorithms, including 3 kinds based on 3D feature, respectively, using singular value decomposition (SVD), orthonormal decomposition (OD) and unit quaternion (UQ), and the fourth based on 2D feature, known as "8-point" algorithm ("8-point"). Results show that those algorithms based on 3D give more precision, more stable and more efficient results than that based on 2D. Among 3 kinds of 3D-based algorithms, SVD is most efficient, meanwhile UQ is a little more stable than else. To "8-point", results show that the nonlinear optimization is effective. To all 4 algorithms, both increasing number of points and decreasing noise level can improve results. All these conclusions are useful to those motion estimation projects.