【作者】 李虹；

【导师】 樊渊； 胡炳彰；

【作者基本信息】 安徽大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 随着工业4.0时代的到来,智能机器人技术的应用得到了广泛关注与研究,同时定位与建图(Simultaneous Localization and Mapping,SLAM)技术能够帮助机器人在未知环境中完成自身定位与地图构建。目前传统视觉SLAM系统大多都基于静态环境的假设,当面对动态场景时,系统性能就会下降。因此,为提高视觉SLAM系统在动态环境下定位与建图的性能,本文针对室内动态环境进行研究,在传统视觉SLAM基础上做出改进,提出了室内动态环境下基于语义融合方法的视觉SLAM算法。本文具体研究工作如下:(1)提出了基于YOLOv4-Tiny(You Only Look Once version 4-Tiny)语义分割的视觉SLAM算法。针对传统视觉SLAM系统在动态环境下鲁棒性差的问题,在ORBSLAM2(Oriented FAST and Rotated BRIEF SLAM version 2)的基础上进行改进,添加一个基于轻量级神经网络YOLOv4-Tiny的语义模块,通过该网络对输入图像中的先验动态对象生成候选框并进行检测剔除,仅保留静态帧用于后续的位姿估计,以提高系统的精度,同时轻量级的神经网络也有效地保证了系统的实时性。(2)提出了融合语义与几何约束的视觉SLAM算法。针对非先验动态对象无法被检测以及由于轻量级网络检测失败导致某些动态对象没有被去除的问题,在语义分割的基础上增加改进的多视图几何模块,在该方法中,图像帧中的动态对象会先经过语义模块的分割,然后再进入几何模块进行检测和剔除。改进的多视图几何法利用关键帧投影的角度差与深度差筛选语义分割所遗漏的动态点,从而进一步提升系统的鲁棒性。(3)提出了基于语义融合的视觉SLAM稠密地图构建算法。针对传统SLAM系统只能生成稀疏点云地图,无法获取场景细节不能满足机器人与环境交互的问题,在上述工作基础上增加了稠密建图线程,实现稠密地图重建,并且动态物体的去除,极大程度地避免了动态对象移动所造成的稠密点云地图残影问题,最终获得了具有更多场景细节与辨识度的稠密点云地图。综上所述,本文以室内动态环境为背景,提出基于语义融合的视觉SLAM算法,有效解决了因动态对象造成的系统定位与建图性能下降的问题,同时生成了具有较高场景辨识度的稠密点云地图,以满足机器人的场景交互需求。更多还原

【Abstract】 With the advent of Industry 4.0 era,the application of intelligent robotics has received extensive attention and research,while Simultaneous Localization and Mapping(SLAM)technology can help robots to complete their own localization and map construction in unknown environments.Currently,traditional vision SLAM systems are based on the assumption of static environments,and the performance of these static environment-based systems degrades when facing dynamic scenes.Therefore,in order to improve the performance of visual SLAM systems for localization and mapping in dynamic environments,this paper conducts research on indoor dynamic environments,makes improvements based on traditional visual SLAM,and proposes a visual SLAM algorithm based on semantic fusion methods in indoor dynamic environments.The specific research work in this paper is as follows:(1)This paper proposes a visual SLAM algorithm based on YOLOv4-Tiny(You Only Look Once version 4-Tiny)semantic segmentation.To address the problem of poor robustness of traditional visual SLAM systems in dynamic environments,we improve on ORB-SLAM2(Oriented FAST and Rotated BRIEF SLAM version 2)by adding a semantic module based on the lightweight neural network YOLOv4-Tiny,which generates candidate frames and detects the a priori dynamic objects in the input image by the network.The network generates candidate frames and detects and rejects dynamic objects in the input image,and only retains static frames for subsequent bit pose estimation to improve the accuracy of the system,while the lightweight neural network effectively ensures the real-time performance of the system.(2)This paper proposes a visual SLAM algorithm that fuses semantic and geometric constraints.To address the problem that non-a priori dynamic objects cannot be detected and some dynamic objects are not removed due to the failure of lightweight network detection,an improved multi-view geometry module is added to the semantic segmentation,in which dynamic objects in image frames are first segmented by the semantic module and then enter the geometry module for detection and rejection.The improved multi-view geometry method uses the angle difference and depth difference of the key frame projection to screen the dynamic points missed by the semantic segmentation,thus further improving the robustness of the system.(3)This paper proposes a semantic fusion-based visual SLAM dense map construction algorithm.For the problem that the traditional SLAM system can only generate sparse point cloud maps and cannot obtain scene details to meet the robot-environment interaction,a dense map building thread is added on the basis of the previous work to achieve dense map reconstruction and dynamic object removal,which greatly avoids the problem of dense point cloud map residuals caused by dynamic object movement and finally obtains a dense point cloud map with more scene details and recognition.The dense point cloud map is finally obtained with more scene details and recognition.In summary,this paper proposes a visual SLAM algorithm based on semantic fusion in the context of indoor dynamic environment,which effectively solves the problem of degradation of system localization and map building performance caused by dynamic objects,while generating a dense point cloud map with high scene recognition to meet the scene interaction requirements of robots.更多还原