Abstract: For traditional SURF (Speeded-Up Robust Feature) algorithm takes a long time for constructing local feature descriptors, an improved SURF algorithm was proposed to meet the real-time requirement. Firstly, the Determinant of Hessian (DoH) matrix was adopted to detect the key points of an image. Non-maximum suppression algorithm and interpolation operation were used to search and position the extreme points. Secondly, gray centroid method was adopted to determine the main direction of the key points. Then a binary descriptor, BRIEF (Binary Robust Independent Elementary Feature), was adopted to describe the key points, and the main direction of the key points was used to construct a directed feature descriptor with the objective of keeping its rotation invariance. Finally, Hamming distance was used to preliminarily determine the matching points. Then, the mismatching points were removed to improve the matching accuracy by ratio detection method and RANSAC (Random Sample Consensus) algorithm. The experimental results show that, when the improved SURF algorithm is applied to the flexible clamping robot, the matching time is reduced from 214.10 ms to 86.29 ms, the matching accuracy is increased by 2.6% compared with traditional SURF algorithm. Therefore, the proposed algorithm can improve the workpiece image matching speed and matching precision of flexible clamping robot effectively.

Key words: visual grasping, Speeded-Up Robust Feature (SURF) algorithm, gray centroid method, BRIEF (Binary Robust Independent Elementary Feature) descriptor, Hamming distance, RANSAC (Random Sample Consensus) algorithm

摘要： 针对传统SURF算法在构建局部特征描述符时耗时较长，无法满足实时性要求的问题，提出了一种改进的SURF算法。首先，运用Hession矩阵行列式（DoH）检测图像中的关键点，并利用非极大值抑制法和插值运算搜索、定位极值点；其次，采用灰度质心法确定关键点的主方向；然后，采用二进制描述符BRIEF对关键点进行描述，并利用关键点的主方向构造带有方向的特征描述符，使其具有旋转不变性；最后，运用汉明距离初步确定匹配点，再用比率检测法和RANSAC算法去除误匹配点，进而获取精准配准。实验结果表明，该改进SURF算法在应用于机器人进行柔性装夹时，对工件图像的平均匹配时间由SURF算法的214.10 ms减少到86.29 ms；而且匹配精度方面比原SURF算法提高了2.6%，因此，所提算法能够有效提高柔性装夹机器人工件图像的匹配速度和匹配精度。

关键词: 视觉抓取, SURF算法, 灰度质心法, BRIEF描述符, 汉明距离, RANSAC算法