基于深度学习的双目立体视觉图像特征点匹配方法  

Deep learning-based feature point matching method for binocular stereo vision images

作　　者：李纪鑫 赫磊 任高明 LI Jixin;HE Lei;REN Gaoming(School of Computer Science and Software,Shaanxi Institute of Technology,xi’an 710300,China)

机构地区：[1]陕西国防工业职业技术学院计算机与软件学院,西安710300

出　　处：《自动化与仪器仪表》2022年第2期57-60,共4页Automation & Instrumentation

基　　金：陕西省教育厅2019年度专项科学研究计划:面向高速网络的流量测量关键技术研究(No.19JK0085)。

摘　　要：针对传统匹配方法分割图像中特征点的权值信息,导致最终匹配过程运行时间较长,针对该问题,设计一种基于深度学习的双目立体视觉图像特征点匹配方法。在选定双目立体视觉图像中的像素点作为处理中心,预处理视觉图像特征,采用随机森林处理方法推断图像中的特征信息,分割图像属性后,利用深度学习方法提取图像特征点,固定处理方向。设定描述梯度,最终实现图像特征点的匹配。随机选定图像数据集作为处理对象,标定图像的特征点后,准备两种传统匹配方法以及设计匹配方法进行实验,结果表明:设计的特征点匹配方法实际所需的运行时间最短。To address the problem that traditional matching methods segment the weight information of feature points in an image, resulting in a long running time for the final matching process, a deep learning-based feature point matching method for binocular stereo vision images is designed. After selecting pixel points in binocular stereo vision images as processing centers, preprocessing visual image features, inferring feature information in images using random forest processing methods, segmenting image attributes, extracting image feature points using deep learning methods, and fixing processing directions. The description gradient is set to finally achieve the matching of image feature points. After randomly selecting the image dataset as the processing object and calibrating the feature points of the image, two traditional matching methods as well as the designed matching method are prepared for experiments, and the results show that the designed feature point matching method actually requires the shortest running time.

关 键 词：深度学习 双目立体视觉 图像特征点 运行时间 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]