结合尺度不变特征的ORB算法改进

Improved ORB Algorithm with Invariant Scale Features

针对ORB(ORiented Brief,方向描述符)算法不具备尺度不变性,且匹配点对错误较多等缺点,结合SURF(Speed-up Robust Features,加速稳健特征)与ORB提出一种新的算法。通过计算积分图像,使用盒子滤波器近似高斯滤波,构建尺度空间,通过Hessian矩阵检测出具备尺度不变的特征点;用ORB对特征点进行描述,采用Hamming距离完成粗匹配;使用改进的RANSAC(RANdom SAmple Consensus,随机抽样一致)减少其迭代次数,同时,去除错误的匹配点对。实验结果表明,在尺度变化时,改进算法的平均准确度为86.3%,约为ORB的3.1倍;综合对比时,改进算法的平均精度可达85.1%,是ORB的2.4倍,平均耗时高于ORB,但远低于SIFT,在不失精度的前提下有效地保证了鲁棒性和实时性。

In view of the fact that the ORB(ORiented Brief) algorithm does not have the scale invariance, and the matching points are more error-prone, a new algorithm combining SURF(Speed-up Robust Features) and ORB is proposed. By calculating the integral image and using the box filter to approximate the Gaussian filter,the scale space was constructed, and the feature points with the same scale were detected by the Hessian matrix. The feature points were described by ORB, and the rough matching was achieved by Hamming distance.The improved RANSAC(RANdom SAmple Consensus) was used to reduce the number of iterations, and at the same time, the wrong matching point pairs were removed. The experimental results show that the average accuracy of the improved algorithm is 86. 3% when the scale is changed, which is about 3.1 times that of ORB. The average accuracy of the improved algorithm is 85. 1%, which is 2. 4 times that of ORB, and the average timeconsuming is higher than that of ORB, but far lower than that of SIFT. The robustness and real-time performance are guaranteed effectively without lossing accuracy.