一种高效的实时视觉姿态测量

Efficient and Real-time Vision-based Pose Estimation

由光学传感器和计算机视觉理论来确定目标物体的姿态参数具有重要的实用价值和理论意义。本文以提高当前物体姿态求解方法的实时性和准确性为目的,提出一种高效的基于特征点的姿态求解方法,该方法首先采用P3P算法求解目标物体的近似姿态,然后将该近似姿态作为初始值嵌入正交迭代(OI)优化算法求取目标姿态的最优值,该方法可以适用于任意特征点数大于3且不共面的情形。实际试验的结果表明,当采用的特征点达到8个时该姿态测量算法的均方根误差保持在3毫弧度(0.17°)以内,而刷新速度达到200帧/秒。

Determining pose parameters for space target based on optical sensors and computer vision theory is full of both practical value and theoretical significance. An efficient point-based pose estimation algorithm is proposed to improve the precision and real-time ability of the pose estimation system. In our method, we first get the initial value for target pose through the classical Perspective-3-Point （P3P） method, then introduce the Orthogonal Iteration （OI） optimization method to make a accurate solutions. This pose estimation approach can be applied in any situation with feature points non-coplanar and the number larger than three. Practical experimental results indicate that when the feature point number is up to eight, the RMS error is less than 3 milliradian （0.17°） and the refresh rate achieves 200 frame/s.