为了求解复杂环境下目标的位姿, 提出了基于多直线对应的加权最小二乘位姿估计算法。首先对模型直线进行等间隔采样, 并沿采样点投影法线方向搜索图像点对应; 然后利用图像点对应局部和全局特性对样本点进行加权; 最后通过优化法向距离实现目标位姿的优化求解。为了解决模型-图像对应错误引起的优化失败问题, 算法在模型-图像点匹配阶段为每个采样点保留多个图像点对应, 通过随机Hough变换(RHT)算法将图像点对应约束在直线上, 并为每条模型直线保留多图像直线对应。在对样本点进行加权时, 综合考虑了样本点自身的属性和样本点同周围点的关系, 有效提高了算法对纹理, 背景, 噪声等的鲁棒性。实验结果表明: 提出的方法能够实现复杂环境下目标位姿的优化求解, 其在x方向、y方向和z方向的角度估计误差分别优于0.4, 0.3和0.1°; 在垂直光轴方向和沿光轴方向的相对位置误差则分别优于0.03%和0.1%。相比单假设方法, 提出的方法能够更有效地克服复杂背景干扰, 实现特殊视图目标位姿的稳定估计。

To estimate the pose of known rigid objects efficiently in a complex environment, a rigid object pose estimation method was proposed by combining multiple line hypothesis and iteratively reweighted least squares. The 1D search was utilized to obtain the corresponding image point along a normal direction for each model sample point by an equal interval sampling. Then, the weight of a sample point was calculated according to the local and global appearances of the corresponding image point for each visible model sample point. The optimized pose parameters were obtained by minimizing the errors between the sample points and their corresponding image points. To avoid the failure of the pose optimization caused by the mismatches of the model and image lines, multiple low level hypotheses were retained for each model sample point in the registration process and they were classified into multiple lines for each potential edge by the Random Hough Transform(RHT). Due to the use of the property of the sample point as well as the relation to the neighbor points, the robustness to disordered background and noise was enhanced in the weighting process. Experiments show the proposed method effectively estimates the poses of freely moving objects in an unconstrained environment. The precisions of the poses on x, y and z axes are better than 0.4°, 0.3° and 0.1° respectively; and those of relative positions perpendicular to the optical axis and along the optical axis are better than 0.03% and 0.1% respectively. Comparisons with the single hypothesis based method demonstrate that the proposed method overcomes the influence induced by the complex background and optimizes the pose parameters in special views.