Abstract: Huge time are needed in feature extraction and matching with classical SIFT in image matching as well as in image retrieval，and a real-time system can not be built when mobile robot located itself with classical SIFT.A modified SIFT is proposed by changing the cycle sequence of sample pickup and setting two different modes on feature pickup.This modified SIFT can be used in omnidirectional vison images.Invariances to image translation，scaling，rotation are kept.A new real-time location system is built with modified SIFT.Examples are presented that this modified SIFT algorithm can realize the localization effectively and stably.

Key words: Scale Invariant Feature Transform（SIFT）, feature extraction, image matching, omnidirectional vision, localization

摘要： 经典SIFT算法的计算量比较巨大，在应用到图像匹配中，尤其是多地图检索的图像匹配定位中时不能满足系统实时性的要求。可用于全景视觉传感器图像的改进SIFT算法，在不改变原算法匹配稳定性的基础上，通过修改原算法的采样规则，同时针对对复杂和简单两种情况下的图像采用不同的采样方式，使系统基本可以达到实时的效果。结果表明，改进算法可以实现高效、准确的定位。

关键词: 尺度不变特征变换, 特征提取, 图像匹配, 全景视觉, 定位