视觉里程计研究综述

视觉里程计(Visual Odometry)作为视觉同步定位与地图构建技术(Visual Simultaneous Localization and Mapping)的一部分,主要通过相机传感器获取一系列拥有时间序列图像的信息,从而预估机器人的姿态信息,建立局部地图,也被称为前端,已经被广泛应用在了多个领域,并取得了丰硕的实际成果,它对于无人驾驶、全自主无人机、虚拟现实和增强现实等方面有着重要意义.本文在介绍经典视觉里程计技术框架模块中的各类算法的基础上,对近年来新颖的视觉里程计技术(VO)的研究和论文进行了总结,按照技术手段不同分为两大类——多传感器融合的视觉里程计(以惯性视觉融合为例)和基于深度学习的视觉里程计.前者通过各传感器之间的优势互补提高VO的精度,后者则是通过和深度学习网络结合改善VO的性能.最后通过比较视觉里程计现有算法,并结合VO面临的挑战展望了视觉里程计的未来发展趋势.

A survey of visual odometry

Visual Odometry (VO),which is an important part of visual simultaneous localization and mapping technology,is mainly used for robot pose estimation and local map building through time series images captured by camera sensors.Known as the front end,VO has been widely used in many fields and achieved fruitful practical results,and it is of great significance for unmanned driving,autonomous drones,virtual reality,and augmented reality,etc.In this paper,we summarize the recent research results on the novel visual odometry technology based on introduction of various algorithms in the framework module of classical VO.According to their technical means,the novel methods are divided into two categories,including VO integrated with multiple sensors (take VIO as an example),and VO based on deep learning.The former improves the accuracy of VO by complementing the advantages of various sensors,while the latter is combined with deep learning network.Finally,the existing algorithms of visual odometry are compared,and the future development trend is forecasted based on the challenges faced by VO.