针对位姿求解过程中存在的解不唯一、选解难和解的精度不高等问题, 提出一种基于点特征的单目视觉位姿测量算法。首先, 根据共面4个特征点之间的位置关系, 分别对平行和相交两种情况进行分析; 其次, 根据特征点的空间坐标、图像坐标和空间位置关系, 推导出世界坐标系中3个坐标轴上的向量变换到摄像机坐标系中的单位向量, 进而求解出物体相对于摄像机的初始位姿; 最后,用LM算法对初始位姿进行优化, 得到最终位姿。实验结果表明: 文中算法的合成误差为0.54 mm; 现有的EPnP算法、两点一线算法和P3P算法的合成误差分别为1.28、1.52和4.26 mm; 文中算法的合成误差分别减小了57.8%、64.4%和87.3%, 优于现有的EPnP算法、两点一线算法和P3P算法。

Aiming at the problems that the solutions were not unique, the correct solution was difficult to select and the accuracy was not high during processing the solution, an monocular vision pose measurement algorithm based on point features was proposed. Firstly, according to the position relationship between the four coplanar feature points, the parallel and the intersection conditions were analyzed respectively; Secondly, according to the spatial coordinates, image coordinates and spatial position relationships of the feature points, the corresponding unit vectors in the camera coordinate of the three coordinate axes in the world coordinate were derived, then the initial pose of the object to the camera was obtained; Finally, the initial pose was optimized with the LM algorithm to obtain the final pose. The experimental results show that the synthesis error of the article algorithm is 0.54 mm, the errors of the synthesis of the existing EPnP algorithm, two-point one-line algorithm and P3P algorithm are 1.28 mm, 1.52 mm and 4.26 mm, respectively. The synthesis error of the article algorithm is reduced by 57.8%, 64.4% and 87.3% respectively. All in all, the article algorithm is superior to the existing EPnP algorithm, two-point one-line algorithm and P3P algorithm.