单目位姿估计是计算机视觉中一个基础而重要的问题,在机器人定位、虚拟现实、图像精密测量等领域应用广泛。在实际应用中,参考点坐标不可避免地含有粗差点,导致估计结果偏离真值,为此,提出自适应加权的稳健正交迭代算法。该算法采用稳健估计方法自动识别粗差点,并赋予其较小权值,以提高算法的稳健性。实验结果表明,稳健正交迭代算法求解精度高、稳健性好,可有效抑制不同个数、不同水平的粗差影响。当20个观测点中存在8个水平为60 pixel的粗差点时,本文解算精度分别比经典正交迭代算法和加权正交迭代算法高2个和1个数量级。

Monocular pose estimation is a basic and important problem in computer vision, being widely used in robot positioning, virtual reality and image precision measurement. In practical application, the coordinates of reference points inevitably contain outliers which may lead to an estimating result far from the true value. Therefore, an adaptive weighted robust orthogonal iteration algorithm is proposed. To improve robustness, this algorithm uses a robust estimation method to find out the outliers and suppress their impaction by allocating them smaller weights. The experiment results show that the proposed algorithm is robust with high accuracy. This algorithm can effectively restrain the influence of outliers with different number and levels. When there are 8 outliers of 60 pixel in 20 reference points, the accuracy of this method is 2 and 1 orders of magnitude higher than that of classical orthogonal iteration algorithm and weighted orthogonal iteration algorithm, respectively.