【作者】 张淼；

【导师】 闫丹凤；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 随着自主移动机器人的普及,同时定位与地图构建(SLAM)技术在机器人定位中变得越来越重要。对于视觉SLAM算法,虽然大多数已经建立了很好的理论框架,但在光照变化、白墙、纹理少等环境中仍存在很多挑战。光照变化带来的地图不可重用问题在机器人在长期定位中不可避免。目前针对光照变化的定位解决方案,大多与深度学习结合。本文的目的是通过优化图像变换模型,提高视觉特征对光照变化的鲁棒性,提升不同光照环境下地图重定位效果。论文提出了一种基于匹配和光度误差的图像变换方法(MPT),并将其无缝集成到基于特征的视觉SLAM框架的预处理阶段。实验表明,提出的图像变换方法对提升不同视觉特征的匹配数量都具有较好的效果。此外,使用ROS封装的图像变换模块能够用于多个视觉SLAM系统,并提升其在不同光照环境下的重定位效果。针对纹理较少的环境中机器人追踪定位失败的情况,本文通过在SLAM过程中使用线特征的方法,增加追踪特征数量,创建更多地图数据,优化追踪定位算法。本文提出了一种由粗到细的点线特征提取匹配算法,在追踪定位的过程中,当点特征匹配数量不足阈值时,进行粗粒度的线特征提取与匹配;在局部地图过程中,插入关键帧之后进行细粒度的线特征提取与匹配。实验结果表明,本文所提算法通过一定的计算耗时,增加了地图中的线特征数量,提升了追踪定位过程的鲁棒性,并满足SLAM系统实时性较高的要求。最后,基于上述算法设计和实现对纹理少和光照变化鲁棒的视觉SLAM系统。基于开源的视觉SLAM框架,增加图像变换模块和地图管理模块,优化了定位模块,实现了一个鲁棒视觉SLAM系统。使用校园场景数据测试评估系统,体现出系统的稳定性和鲁棒性。更多还原

【Abstract】 With the popularity of autonomous mobile robots,simultaneous localization and mapping(SLAM)technology in robot localization is becoming more and more important.For visual SLAM algorithms,although most of them have established a good theoretical framework,there are still many challenges in environments such as illumination changes,white walls,and few textures.The problem of map non-reusability caused by illumination change is an inevitable problem for robots in long-term localization.Most of the current localization solutions for illumination changes are combined with deep learning.The purpose of this thesis is to improve the robustness of visual features to illumination changes by optimizing the image transformation model,and to improve the effect of map re-localization in different lighting environments.This thesis proposes an image transformation method based on matching and photometric error(MPT)and integrates it seamlessly into the preprocessing stage of a feature-based visual SLAM framework.Experiments show that the proposed image transformation method has a good effect on improving the matching number of different visual features.In addition,the image transformation module encapsulated in robot operating system(ROS)can be used with multiple visual SLAM systems to improve its re-localization effect in different lighting environments.Faced with the failure of robot tracking and localization in an environment with less texture,this thesis increases the number of tracking features,creates more map data,and optimizes the tracking and localization algorithm by using line features in the SLAM.This thesis proposes a feature extraction and matching algorithm from coarse to fine lines.In the process of tracking and localization,when the matching number of point features is less than the threshold,coarse-grained line feature extraction and matching are performed.In the local mapping process,fine-grained line feature extraction matching is performed after inserting keyframes.The experimental results show that the algorithm increases the number of line features in the map through a certain timeconsuming calculation,which improves the robustness of the tracking and localization process and meets the high real-time requirements of the SLAM system.Finally,a robust visual SLAM system with less texture and illumination changes is designed and implemented based on the algorithm above.On the basis of the open-source visual SLAM framework,the system adds an image transformation module and a map management module,optimizes the localization module,and finally integrates to implement a robust visual SLAM system.The integrated system is tested and evaluated using the campus scene data,which reflects the stability and robustness of the system.更多还原