改进ORB特征提取算法的ORB-SLAM2定位研究

Research on ORB-SLAM2 Localization Based on Improved ORB Feature Extraction Algorithm

针对ORB(oriented fast and rotated brief)算法提取图片特征时存在的特征点分布不均匀问题,提出了一种改进的ORB算法。通过密度峰值聚类算法计算四叉树分割后的各区域特征点密度峰值,并对密度峰值高的区域使用四叉树多次分割;同时在初始阈值(iniThFAST)和最小阈值(minThFAST)中加入中间阈值(middleThFAST),提高特征点灰度值差值,减少冗余特征点提取,提高图片特征点均匀性。对特征点提取与匹配进行了实验验证,结果表明改进算法对比传统ORB算法在均匀度和匹配效率方面均有明显提升。结合ORB-SLAM2定位实验表明,改进算法的定位精度平均提高16.6%,每帧追踪时间平均减少12.4%,有效保证实际定位过程的精确性和实时性。

Aiming at the uneven distribution of feature points in image feature extraction by ORB(oriented fast and rotated brief)algorithm,an improved ORB algorithm is proposed.The peak density of feature points in each region after quadtree segmentation is calculated by density peak clustering algorithm,and the region with high density peak is segmented by quadtree multiple times.At the same time,the intermediate threshold(middleThFAST)is added to the initial threshold(iniThFAST)and minimum threshold(minThFAST)to improve the difference of the gray value of feature points,reduce the extraction of redundant feature points,and improve the uniformity of feature points.The experimental verification of feature point extraction and matching shows that the improved algorithm has obvious improvement in uniformity and matching efficiency compared with the traditional ORB algorithm.Combined with the positioning experiment of ORB-SLAM2,the positioning accuracy of the improved algorithm is improved by 16.6%on average,and the tracking time of each frame is reduced by 12.4%on average,which effectively ensures the accuracy and real-time performance of the actual positioning process.