约束可满足方法对单幅图片的三维猜测与个体识别  

3D Explanation of Photo and Object Recognition Using the Constraint Satisfaction Method

作　　者：危辉[1] 张维[1] 

机构地区：[1]复旦大学计算机科学系认知算法模型实验室,上海200433

出　　处：《小型微型计算机系统》2010年第2期338-343,共6页Journal of Chinese Computer Systems

基　　金：上海市科技发展软科学研究基金项目(066921049)资助

摘　　要：从二维图像提取到的特征点相对位置信息,利用对于目标物体的先验知识如边缘比例、夹角特征等,形成基于投影成像规则的变量约束方程,从而将物体姿态计算问题转化为一个非线性优化问题,通过多个局部约束线索形成全局约束,能够迅速对物体的姿态做出比较准确的估计,从而推算出视角等信息.实现这个转化的一个前提假设是,对于我们理解图像中物体的姿态,特征位置点的相对位置而不是绝对位置,起着关键作用.因此计算是在一个假设的深度上进行的,从效果来看这样的假设并不影响物体的位姿计算.本文的方法计算量小,利用几何特征来识别稳定、可靠、泛化能力好,实践证明使用几何特征的约束可满足方法能获得关于姿态的极少量可能解,识别出的姿态在各条边之间的比例关系上具有不变性,继而可以将其应用于不变性识别的实际问题之中.The position information about target points extracted from an image can be formed into constraint equations based on the perspective projection rules and priori knowledge about object such as angles and ratio of edges.Then the problem of pose estimation can be converted into a nonlinear optimization problem.The overall solution of pose estimation can be gotten from local clues.The observation perspective can also be estimated from this information.During the calculation procedure it is assumed that the object is located in some depth and the relative depth not absolute depth of different points on object plays the important role for the problem of pose and observation perspective estimation.The experiments show that the error caused by this assumption can be tolerated for the problem of pose estimation.This approach needs a small amount of calculation.The use of the geometric characteristics makes the process of calculation stable,reliable,and with fine generalization ability.The method of constraint satisfaction can generate only a very small amount of possible solutions.The relationship within edges in the pose of estimation maintains invariant.This can be used for the problem of invariant recognition.

关 键 词：约束可满足 PNP问题 三维解释 不变性识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]