摘要
目标位姿追踪技术是指通过图像信息,解算物体坐标系相对于相机坐标系的旋转矩阵以及平移向量。该领域近年来发展迅速,基于特征点匹配以及姿态解算的位姿追踪方法在精度以及算法速度上已基本满足要求。但对于表面纹理贫乏的目标,由于无法对其进行特征点提取并匹配,导致算法失效。本文使用物体模型以及场景灰度图像构建方向倒角距离目标函数,将追踪问题转化为寻优问题。在现有算法基础上,结合物体位姿变换与图像坐标改变的雅可比矩阵,使用解析偏导数替换数值偏导数。针对追踪过程中出现的背景扰动问题,使用场景梯度方向与模型边缘方向计算模型点优化权重,及时降低误匹配或干扰点对姿态解算的影响,提高复杂情况下的追踪精度。测试阶段,在渲染视频测试中,平均角度误差在1.30°以内,平均平移向量误差在0.85毫米内;且在难度较高的Rigid Pose数据集中也取得了不错的效果。测试结果表明该算法在复杂条件下仍能实现高精度追踪。
We propose a novel model-based method for estimation and tracking the six degrees of freedom(6DOF)pose of rigid objects of arbitrary shapes.Relevant researches have been developing rapidly in recent years,and the pose tracking method based on feature point matching and perspective-n-point algorithm has basically met the requirements on accuracy and speed.However,in industrial applications and augmented reality,most targets are textureless objects.It is difficult to extract feature points from the surface of such objects,which leads to large error and even failure of the method based on feature points.In this paper,object model and scene gray image are used to construct the object function by direction chamfer distance,and the tracking problem is transformed into the optimization problem.Based on the existing work,our approach constructs the Jacobian matrix between the three-dimensional point space coordinate and the coordinate of the corresponding points on the image,to replace the existing numerical derivative optimization method.In order to solve the occlusion problem,we use the picture gradient direction and the object edge direction to judge whether the point is occluded,and calculate the optimal weight by the difference between the two direction.When the object is occluded,the tracking algorithm can automatically reduce the influence of occluded points on the pose optimization,making the registration process more inclined to the unoccluded points and improving the tracking accuracy in this case.we leverage CG render video,real gray-scale video and open source dataset to evaluate our tracking algorithm.In the Rigid Pose dataset test,the average angle error of our algorithm is 1.835°,andthe a verage translation vector error is 0.989 mm,better than the existing methods.Combining the analytic derivatives and the optimal weight approach,the accuracy of the registration is higher than the existing methods.Since our methods have a lower computational complexity,lower device requirements.
作者
陈策
岳继光
董延超
沈润杰
CHEN Ce;YUE Ji-guang;DONG Yan-chao;SHEN Run-jie(College of Electronic and Information Engineering,Tongji University,Shanghai 201804)
出处
《新一代信息技术》
2018年第5期1-10,共10页
New Generation of Information Technology
基金
国家自然科学基金项目(61873189)
上海市自然科学基(18ZR1442500)
中央高校基本科研业务费(KX0080020172601)。
关键词
位姿追踪
工业视觉
模型匹配
解析寻优
Pose tracking
Industrial vision
Model matching
Analytic optimization
