结合YOLO的ORB双目图像匹配方法研究

针对双目视觉中ORB(Oriented Fast and Rotated Brief)图像匹配算法准确率不高以及会出现不同物体之间特征点错误匹配的问题,提出一种将YOLO(You Only Look Once)目标检测算法和ORB算法结合的双目图像匹配方法.该方法首先使用YOLO的卷积网络提取图像特征,并采用多尺度预测目标区域坐标和类别信息,以解决小目标与多目标识别不准的问题;接着,使用FAST(Features from Accelerated Segment Test)算子检测特征点和BRIEF(Binary Robust Independent Elementary Features)算子描述特征点,并利用ORB算法进行粗匹配;最后用去误匹配算法判断并去除不同类别和位置信息目标框中的匹配点.实验结果表明,该方法在单目标、双目标和多目标双目图像中的匹配准确率相较传统ORB匹配算法精度都有所提升.

Research on ORB Binocular Image Matching Method Based on YOLO

Aiming at the problem that the accuracy rate of ORB image matching algorithm in binocular vision is not high enough and feature points of different objects are mismatched,this paper presents binocular image matching method that combines YOLO targetdetection algorithm with ORB algorithm.Firstly,to solve the problem that the recognition to small targets and multiple targets is inaccurate,this method uses YOLO convolutional neural network to extract image features,and adopts the method of multiple scale to predict coordinates and class information of target regions.Secondly,FAST operator is used to detect feature points while BRIEF operator is used to describe feature points.What’s more,ORB algorithm is applied for rough matching.Lastly,matching points in the target boxes with different labels or in different locations are judged and removed by the use of removing mis-matching algorithm.According to experimental results,compared with traditional ORB matching algorithm,the matching accuracy rates of this method in single-target,double-target and multi-target binocular image are improved.