单目视觉SLAM综述  

Review of Monocular Visual SLAM

作　　者：郑义桀 罗健欣 陈卫卫 潘志松 孙海迅 ZHENG Yi-jie;LUO Jian-xin;CHEN Wei-wei;PAN Zhi-song;SUN Hai-xun(College of Command and Control Engineering,Army Engineering University of PLA,Nanjing 210014,China)

机构地区：[1]陆军工程大学指挥控制工程学院,江苏南京210014

出　　处：《软件导刊》2022年第12期242-251,共10页Software Guide

基　　金：国家自然科学基金项目(62076251)。

摘　　要：单目视觉SLAM因单目相机体积小、功耗低、信息获取丰富受到了广泛使用。为了深入分析单目视觉SLAM的优势,首先,简述单目视觉SLAM的基本原理,从特征点检测、相机姿态估计、选取关键帧、创建地图、地图及相机姿态优化、闭环检测等方面总结SLAM的关键环节技术。然后,基于特征方法、直接法和混合半直接法对当前主流单目视觉SLAM系统的设计框架进行详细介绍,分析20余种较为流行的系统性能特点和适用场景。最后,介绍深度学习在相机姿态估计、创建地图、闭环检测等环节中的应用,并与传统方法进行比较,以讨论单目视觉SLAM的发展趋势。Monocular vision SLAM is widely used because of its small size, low power consumption and rich information acquisition. In order to analyze the advantages of monocular vision SLAM in depth. Firstly, the basic principle of SLAM based on monocular vision is briefly introduced, and the key technologies of SLAM are summarized from the aspects of feature point detection, camera pose estimation, key frame selection, map creation, map and camera pose optimization, and closed-loop detection. Then, based on the feature method, direct method and hybrid semi direct method, the design framework of the current mainstream monocular vision slam system is introduced in detail, and more than 20 popular system performance characteristics and applicable scenarios are analyzed. Finally, the application of deep learning in camera attitude estimation, map creation and closed-loop detection is introduced, and compared with traditional methods to discuss the development trend of monocular vision SLAM.

关 键 词：计算机视觉 单目视觉 SLAM 同步定位与地图创建 视觉里程计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]