摘要
图像拼接的关键技术在于特征点的匹配。Lowe等提出的基于SIFT特征算子的匹配算法具有尺度、旋转、视角、光照不变性,能够有效地用于目标的三维重建以及复杂目标识别。该算子使用128维向量来表示每个特征点,使处理的数据量较大,难以满足实时性的要求。本文通过改变特征点描述子的结构实现了特征向量的简化,并且提出基于视差梯度约束的特征点匹配算法,在匹配过程中使用最小中值估计算法去除伪匹配点对。实验说明,当图像存在较大的变形、畸变和噪声影响时的情况下,在保证算法的鲁棒性同时,能够降低图像匹配的计算量,从而保证算法的实时性。
The matching of feature points is the key technique of image mosaic. SIFT operator proposed by Lowe, et al, with scale, rotation, perspective, illumination invariance, can be effectively used to re- construct a 3D object and recognize a complex object. This operator denotes each feature point by 128-di- mensional vector, so large amount of data needs to be processed and the real-time requirement can not be achieved. By changing description structure of feature point, the number of dimension of feature vector is reduced in this study. In addition, a matching algorithm based on the disparity gradient constraint is per- formed. In the process of matching, the algorithm uses Least-Median-Squares estimation to eliminate false matching pairs. The results of experiments demonstrate that this algorithm is able to reduce the matching time as well as has good robustness in the circumstances with more deformation, distortion and noise. Therefore, based on this algorithm the real-time requirement can be achieved.
出处
《中国体视学与图像分析》
2009年第2期156-161,共6页
Chinese Journal of Stereology and Image Analysis