基于深度学习图像特征匹配的双目测距方法  

Binocular Distance Measurement Method Based on Image Feature Matching Using Deep Learning

作　　者：姚尧 张生[1] YAO Yao;ZHANG Sheng(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《软件导刊》2022年第3期207-212,共6页Software Guide

基　　金：上海市科研技术委员会科研计划项目(19511105103)。

摘　　要：针对现有双目视觉测距方法中存在的精度低、抗干扰能力弱、鲁棒性差等问题,提出一种基于深度学习图像特征匹配的双目深度测距方法。首先将双目图像通过一个自监督训练的特征提取网络,通过两个解码器获取双目图像的特征点与描述符,然后根据描述符进行特征点的匹配以获取视差,最后根据双目相机的相似三角形原理获取目标的真实深度。在KITTI数据集上的验证实验结果表明,建立的特征提取网络性能优于传统方法与其他基于深度学习的方法。双目测距实验结果表明,该方法在20~150cm范围内的平均误差为0.58%,满足精确测距要求。To solve the problems of low precision,weak anti-interference ability and poor robustness in existing binocular distance measurement method,a binocular distance measurement method based on image feature matching using deep learning is proposed.Firstly,images from the binocular camera go through a self-supervised keypoint detection network.The keypoints and descriptors can be obtained by two decoders.Then,the keypoints are matched with the descriptors and compute disparity.Finally,true distance is obtained according to the principle of similar triangles of binocular camera.Experimental results on KITTI datasets show that the proposed feature detection network demonstrates better performance than traditional methods and other methods based on deep learning.Experimental results on binocular distance measurement show that the average error of the proposed method is 0.58%within the range of 20~150cm,which means the proposed method can meet the requirements of accurate ranging.

关 键 词：双目测距 深度学习 特征匹配 自监督训练 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]