针对现有Harris与尺度空间结合算法存在复杂度高、准确度低、实时性差的问题，提出一种改进算法。建立尺度不变特征转换（SIFT）图像金字塔，在其上检测Harris特征点，以32维向量作为特征描述符。利用向量相似性匹配特征点。改进了经典K-means算法，不设固定初始值，以距离大、相关性小的候选初始类中心点为初始类中心点，特征点归入距离最小类中。从2幅图像的各类特征点中随机抽取3对匹配点，组成一对三角形，利用三角形相似性进一步筛选匹配点。令改进的RANSAC算法根据匹配点误差距离，分别为所有的匹配点赋值，令所有的匹配点共同评估变换模型，选出最优解。实验结果表明，该算法比SIFT算法、Harris算法提取的特征点减少了约22%，匹配正确率提高了约13%，运算时间减少了约4.7%。

For the existing Harris scale space combined algorithms with high complexity， low accuracy and poor real-time performance， an improved algorithm is proposed. Establish a scale space according to scale-invariant feature transform（SIFT） algorithm to detect Harris feature points， describe features using a 32 dimensional vector. Use vector similarity to match feature points. The Classic K-means algorithm is improved. It has not a fixed initial value， takes the candidate class center point with large distance and low correlation as the initial class center point and categorizes feature points into the class with the smallest distance. Three pairs of matching points were randomly selected from classes of feature points of two images to form a pair of triangles. The matching points are further filtered by triangles similarity.The improved RANSAC algorithm assigns values to all match points based on the absolute values of match point errors to jointly evaluate the transformation model. The experimental results show that the number of feature points extracted by this algorithm is about 22% less than that of SIFT and Harris algorithm， the matching accuracy is improved by about 13%，and the operation time is decreased by about 4.7%.