摘要
为了实时测量非合作纹理运动目标相对位姿,提出一种单目视觉同步建模与位姿计算方法。选择具有良好特征共视关系与分布的模型帧,增量式恢复目标三维共视特征模型,实现非合作目标的合作化,并依靠运动预测模型实现基于特征跟踪的位姿计算。根据模型对应网格的拓扑关系估计目标表面未知区域特征三维信息,结合网格面法向场进行局部光束平差优化,同时利用闭环优化抑制尺度漂移,在减少特征模型恢复误差的同时提高位姿计算精度。实验结果表明,在非结构化环境中能够在线恢复目标三维信息,实现目标相对位姿准确计算,特征平均重投影误差小于1.5 pixel,位姿计算平均绝对误差为4.29 mm和1.54°,平均计算时间小于120 ms,为基于单目视觉的在线三维感知与测量建模提供技术支撑。
To measure the relative pose of moving non-cooperation textured objects in real time,a monocular simultaneous modeling and pose calculation method was proposed.A 3D covisibility model was incrementally constructed with frames containing the highest covisibility of features and best distribution to achieve cooperation between non-cooperation objects.Subsequently,the relative pose of the object was calculated via feature tracking by the motion prediction model.The mesh of the model was used to restore the 3D information of feature points that were distributed in an unknown area of the object surface.To reduce model error and improve the accuracy of pose estimation,bundle adjustment optimization was performed using a facet normal field,and the scale drift was reduced using closed-loop optimization.Experiments show that the method isa real-time online system that can recover 3D information of object in unstructured environments and accurately estimate relative poses in unstructured environments to provide technical support for 3D sensing and measurement modeling based on monocular vision.The mean reprojection error(MRE)of the features using the proposed method is less than 1.5 pixels,and the mean absolute error(MAE)of pose calculation is 4.29 mm and 1.54°while the average time consumption is less than 120 ms.
作者
冯肖维
谢安安
肖健梅
王锡淮
FENG Xiao-wei;Xie An-an;Xiao Jian-mei;Wang Xi-huai(Department of Electrical Automation , Shanghai Maritime University, Shanghai 201306, China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2020年第8期1775-1784,共10页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.61503241,No.61801287)。
关键词
单目视觉
非合作目标
位姿计算
特征建模
非线性优化
monocular vision
non-cooperation objects
pose calculation
feature modelling
nonlinear optimization
