从单幅灰度图像中求解三维物体姿态的方法  

Algorithm for Constrained 3D Object Pose Determination from Single Intensity Images

作　　者：成兰[1] 邢哲[2] 田原[3] 

机构地区：[1]安阳工学院科研处,安阳455000 [2]山东科技大学电气工程系,泰安271000 [3]中国科学院自动化研究所综合信息系统研究中心,北京100080

出　　处：《模式识别与人工智能》2007年第3期428-434,共7页Pattern Recognition and Artificial Intelligence

摘　　要：诸如车辆之类的三维物体的运动常常被约束于一个平面上.这种平面约束显著减少二维图像到三维目标的匹配自由度,从而使得这类三维目标的识别问题得到极大简化.基于此,本文提出一种从单幅定标的灰度图像中求解约束于一已知平面的三维目标的方法.该方法首先计算出给定模型所有可能的旋转角度,再通过聚类获得一些可能的旋转角度类.在每个角度类中,我们对目标所有可能的位置进行计算,并通过聚类获得在此旋转角度下的位置类.之后将候选的角度和位置进行过滤并进行优化以获得最终结果.该算法对于三维目标的部分遮挡具有较好的适应性,其有效性得到实验验证.Objects, such as road vehicles, are often constrained on a ground plane. The ground-plane constraint significantly reduces the pose redundancy of 2D image and 3D line matching. A method is proposed to locate 3D objects on a known plane from a single calibrated intensity image. Firstly, all possible rotation angles of a given model are computed and clustered to obtain some of the most probable rotation angles. Then all possible positions are calculated and clustered for each rotation angle. The candidate rotation angles and positions are filtered and refined to obtain the final results. The proposed method is capable of coping with moderate occlusion. The effectiveness of the method has been verified by experiments.

关 键 词：3D物体识别 基于模型的视觉 聚类 姿态求解 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]