第 35 卷 第 1 期
2015 年 1 月
Vol. 35, No. 1
January, 2015
光 学 学 报
ACTA OPTICA SINICA
0115003-
基于多级直线表述和 M-估计的三维目标位姿跟踪优
化算法
张跃强
1, 2
苏 昂
1, 2
刘海波
1, 2
尚 洋
1, 2
于起峰
1, 2
1
国防科技大学航天科学与工程学院, 湖南 长沙 410073
2
湖南省图像测量与视觉导航重点实验室, 湖南 长沙 410073
摘要 为了实现复杂环境下已知模型目标姿态的快速跟踪和估计,提出了一种结合三维(3D)粒子滤波跟踪和 M-估
计优化的位姿跟踪估计算法。基于直线的多级向量表示构造了新颖的模型直线和图像直线相似性度量函数;基于
粒子滤波跟踪的姿态设计了模型直线和图像直线快速对应方法;利用 M-估计实现了目标姿态的优化估计;利用重
要性 采样方法 将优化姿态有效 地融合到 了粒子 滤波框架 。另外根据预测 的目标位 姿定义 了图像动 态感兴趣区域
(ROI),极大地减少了特征检测和搜索的时间。实验表明,所提方法能够实现复杂环境下 自由移动目 标的快速跟 踪
和位姿的高精度解算,相比已有方法,所提方法在跟踪精度,计算效率以及稳健性上均有优势。
关键词 机器视觉;三维跟踪;直线表达;M-估计;粒子滤波
中图分类号 TP391.4; TP242.6 文献标志码 A
doi: 10.3788/AOS201535.0115003
Three Dimensional Rigid Objects Pose Tracking and Optimization
Based on Multilevel Line Representation and M-Estimation
Zhang Yueqiang
1,2
Su Ang
1,2
Liu Haibo
1,2
Shang Yang
1,2
Yu Qifeng
1,2
1
College of Aerospace Science and Engineering, National University of Defense Technology, Changsha, Hunan 410073, China
2
Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation, Changsha, Hunan 410073, China
Abstract To track and estimate the pose and position of known rigid objects efficiently in complex
environment, a method coupled three dimensional (3D) particle filter (PF) framework with M- estimation
optimization in a closed loop is proposed. A novel similarity observation model is constructed based on
multilevel line representation; line correspondences between 3D model edges and two dimensional (2D) image
line segments are received easily based on the tracking state of PF. After that, line correspondences are
provided for M- estimation to optimize the pose and position of objects. The optimized particles are fused into
the particle filter framework according to the importance sampling theory. Moreover, to speed up the proposed
method, line detection and search space is limited in a local region of interest (ROI) predicted by PF.
Experiments show that the proposed method can effectively track and accurately estimate the pose of freely
moving objects in unconstrained environment. Comparisons on synthetic and real images demonstrate that
proposed method greatly outperforms the state-of-art method in accuracy and efficiency.
Key words machine vision; three dimensional tracking; line representation; M-estimation; particle filter
OCIS codes 150.1135; 330.5000; 150.5758; 110.4153
收稿日期: 2014-06-03; 收到修改稿日期: 2014-08-14
基金项目: 国家 973 计划(2013CB733100)
作者简介: 张跃强(1987—),男,博士研究生,主要从事图像测量、计算机视觉相关方面的研究。E-mail: zyoungnudt@yahoo.com
导师简介:于起峰(1958—),男,教授,博士生导师,中国科学院院士,主要从事空天图像测量与视觉导航方面的研究。
E-mail: yuqifeng@vip.163.com
1 引 言
三维(3D)跟踪的目的是连续恢复出相机相对于场景的位置和旋转,或者目标相对于相机的相对运动,可
1