一种基于区域密度划分改进的ORB特征提取算法

An Improved ORB Feature Extraction Algorithm Based on Region Density Division

摘要: 原ORB算法在面对特征点分布不均匀的图像做特征提取时，常会出现特征点提取效率低，特征点冗余的问题，导致特征点在匹配时出现误匹配，甚至位姿丢失的情况。针对这一问题，提出了一种基于区域密度划分改进的ORB特征提取算法。首先对图像构造图像金字塔，确保图像的尺度不变性。其次用FAST特征点提取算法对图像进行特征点提取，提取完毕后，使用核密度估计方法，以每个特征点为中心，计算该点周围的密度值，并采用聚类算法将密度分布相似的特征点看成一簇，根据每一簇的特征点密度和图像平均密度的比值，将图像区域划分为三类，即特征点密集区域，特征点稀疏区域和特征点均匀区域。然后利用自适应四叉树法对三个区域图像进行分割，并根据每个区域的特征点密度计算阈值，达到特征点筛选均匀的目的。然后利用自适应非极大值抑制法对特征点进行最佳筛选，并使用BRIEF算法计算出特征点的描述子。最后进行特征点匹配。实验结果表明，本文算法相较于传统的ORB算法有效地减少了冗余特征点的数量，降低了特征点的重叠率，在特征匹配的精度和效率上有了明显的提升。

Abstract: When the original ORB algorithm is used for feature extraction in the face of images with uneven feature point distribution, the problems of low feature point extraction efficiency and redundancy of feature points often occur, resulting in mismatching of feature points and even loss of pose. To solve this problem, an improved ORB feature extraction algorithm based on region density division was proposed. Firstly, the image pyramid is constructed to ensure the scale invariance of the image. Secondly, the FAST feature point extraction algorithm is used to extract the feature points of the image. After the extraction, the kernel density estimation method is used to calculate the density values around the point with each feature point as the center, and the clustering algorithm is used to treat the feature points with similar density distribution as a cluster. According to the ratio of the feature point density of each cluster to the average image density, the image area is divided into three categories. That is, feature point dense region, feature point sparse region and feature point uniform region. Then the adaptive quadtree method is used to segment the image of three regions, and the threshold is calculated according to the density of feature points in each region, so as to achieve the purpose of uniform feature point screening. Then, adaptive non-maximum suppression method was used to optimize the feature points, and BRIEF algorithm was used to calculate the de-scriptor of the feature points. Finally, feature point matching is carried out. Experimental results show that compared with the traditional ORB algorithm, the proposed algorithm can effectively re-duce the number of redundant feature points, reduce the overlap rate of feature points, and signifi-cantly improve the accuracy and efficiency of feature matching.