【作者】 张翔；

【导师】 李磊；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 视觉SLAM(Visual Simultaneous Localization and Mapping)是指通过视觉传感器获取环境信息,并通过算法对这些信息进行处理,同时实现对设备自身位置的估计和对环境地图的构建。视觉SLAM技术在自动驾驶、机器人导航、虚拟现实等领域有着广泛的应用。然而,在实际应用中,由于环境复杂多变,相机并非匀速运动等问题,视觉SLAM系统的鲁棒性和可靠性仍然存在诸多挑战。因此,对视觉SLAM系统的鲁棒性提升的研究具有重要的意义。视觉SLAM系统的核心是数据关联和优化。本文将围绕数据关联和优化,根据当前视觉SLAM系统在不同工作场景下存在的鲁棒性问题,从三个方面对视觉SLAM系统鲁棒性提升问题进行研究:(1)高纹理静态场景下视觉SLAM鲁棒性提升的研究;(2)低纹理静态场景下视觉SLAM系统鲁棒性提升的研究;(3)动态场景下视觉SLAM系统鲁棒性提升的研究。基于以上三个方面的研究方向,本文较为系统地分析与完善了当前针对视觉SLAM鲁棒性提升研究的不足之处,据此,本文具体展开了以下研究工作:1、对于高纹理静态场景下视觉SLAM鲁棒性提升的问题,本文围绕相机运动模型和大动态范围场景下特征点提取和匹配问题展开研究。针对恒速运动模型不能很好地刻画相机实际运动过程的问题,本文首先从理论层面分析了利用恒速运动模型进行数据关联时存在的误差,然后基于误差分析结果提出了恒加速运动模型。在TUM数据集上的实验结果表明,本文提出的恒加速运动模型能够有效提升系统的精度。其中在有回环的测试序列上,在绝对轨迹误差(ATE)和相对位姿误差(RPE)的均方根误差指标上,采用恒加速运动模型的视觉SLAM系统的轨迹估计结果相较于采用恒速运动模型的视觉SLAM系统的轨迹估计结果,精度提升了5.0%,在无回环的测试序列上精度分别提升了26.7%、23.9%。针对大动态范围场景下图像特征点提取和匹配容易失败的问题,本文基于Retinex图像分解理论,采用图像照度均衡算法均衡图像的亮度信息。实验结果表明,对于图像中某些暗部区域,经过照度均衡算法处理后的图像更容易完成特征点的提取和匹配。2、对于低纹理场景下基于图像特征点的视觉SLAM鲁棒性急剧下降的问题,本文引入了直线特征和平面特征。其中针对传统平面匹配算法存在匹配速度慢、匹配步骤复杂的问题,本文提出了快速平面匹配算法。该算法单步平面匹配操作的时间复杂度为O(1),整体时间复杂度为O(n)。实验结果表明,本文提出的快速平面匹配算法相较于传统平面匹配算法,单步匹配操作平均时间消耗和最大时间消耗小接近两个数量级,方差小接近四个数量级。从建图效果看,本文提出的平面匹配算法精度更高。本文基于点线面特征提出了多特征融合的定位和建图方案,实验结果表明,本文提出的算法能有效降低系统对图像点特征的依赖,并且在TUM数据集有结构低纹理的测试序列上能够正常工作。考虑到视觉SLAM系统容易因累积误差导致系统估计结果出现漂移现象,本文采用了基于曼哈顿坐标系的相机位姿低漂移估计方案。通过分析三维空间中相机位姿变换的代数学关系,本文进一步提出了基于三维空间中非正交坐标系的相机位姿低漂移估计方案。在都关闭回环检测功能的前提下,测试结果表明,本文提出的相机位姿低漂移估计方案相较于原始系统在绝对轨迹误差的平均值指标上,精度提升了7.1%;在绝对轨迹误差的均方根误差指标上,精度提升了12.5%。3、针对动态场景下视觉SLAM系统容易受运动物体干扰的问题,本文采用语义分割神经网络提取场景中会主动运动的物体区域,然后进一步提取该区域的外轮廓。本文根据特征点到外轮廓的距离,将图像中提取到的特征点划分为不同运动特性的特征点集。然后基于ICP算法在先验静止特征点集上对相机位姿进行初始化,然后利用这个粗略的初始化位姿结合场景流几何约束以及对极几何约束,对先验被动运动特征点集进行运动一致性检测,最终得到静止特征点集合,后续将基于静止特征点集合进行相机位姿的精确跟踪。实验结果表明,本文提出的方案在TUM数据集的动态场景测试序列上的测试结果相较于原始系统的测试结果在绝对位姿误差的均方根误差指标上,精度提升了84.5%-98.4%;在相对位姿误差的均方根误差指标上,精度提升了24.0%-65.2%。更多还原

【Abstract】 Visual SLAM(Visual Simultaneous Localization and Mapping)refers to the acquisition of environmental information through visual sensors and the processing of this information through algorithms,while simultaneously estimating the device’s own position and constructing an environmental map.Visual SLAM technology has a wide range of applications in fields such as autonomous driving,robot navigation,and virtual reality.However,in practical applications,there are still many challenges to improving the robustness and reliability of visual SLAM systems,such as the complexity and variability of the environment and the non-uniform motion of the camera.Therefore,research on improving the robustness of visual SLAM systems is of great significance.The core of visual SLAM system is data association and optimization.This thesis focused on the issues of robustness of visual SLAM systems in different working scenarios and conducted research on improving the robustness of visual SLAM systems from three aspects:(1)research on improving the robustness of visual SLAM in high-texture static scenes;(2)research on improving the robustness of visual SLAM in low-texture static scenes;(3)research on improving the robustness of visual SLAM in dynamic scenes.Based on the above research directions,this thesis systematically analyzed and improved the shortcomings of current research on improving the robustness of visual SLAM.Accordingly,this thesis carried out the following research work:1.This thesis focused on improving the robustness of visual SLAM in high-texture static scenes by addressing the problems of camera motion modeling and feature point extraction and matching in large dynamic range scenes.In order to address the issue that the constant velocity motion model could not well characterize the actual camera motion,this thesis first analyzed the errors in data association using the constant velocity motion model from a theoretical perspective,and then proposed a constant acceleration motion model based on the analysis results.The experimental results on the TUM dataset showed that the proposed constant acceleration motion model could effectively improve the system’s accuracy.Specifically,on the test sequence with loop closures,the trajectory estimation accuracy of the visual SLAM system using the constant acceleration motion model was improved by 5.0% for the root mean square error(RMSE)of absolute trajectory error(ATE),compared to that using the constant velocity motion model.On the test sequence without loop closures,the accuracy was improved by 26.7% and 23.9%.To address the problem of feature point extraction and matching in large dynamic range scenes,this thesis adopted the Retinex image decomposition theory and the image illumination equalization algorithm to equalize the brightness information of the image.The experimental results showed that the image after processing by the illumination equalization algorithm was more likely to complete feature point extraction and matching in some dark areas.2.This thesis addressed the problem of the drastic decline in the robustness of visual SLAM based on image feature points in low-texture scenes by introducing line features and plane features.To address the issues of slow matching speed and complex matching steps in traditional plane matching algorithms,this thesis proposed a fast plane matching algorithm.The time complexity of each step in the proposed algorithm was O(1),and the overall time complexity was O(n).Experimental results showed that the proposed fast plane matching algorithm had significantly lower average and maximum time consumption for each step than traditional algorithms,with much smaller variance.Moreover,the proposed algorithm achieved higher accuracy in map building.Based on point,line,and plane features,this thesis proposed a multi-feature fusion positioning and mapping solution,which effectively reduced the system’s dependence on image point features.Experimental results on the TUM dataset with structureless low-texture test sequences demonstrated the effectiveness of the proposed algorithm.Considering that visual SLAM systems are prone to drift due to cumulative errors,this thesis adopted a low-drift camera pose estimation solution based on a Manhattan coordinate system.Furthermore,based on the algebraic relationship of camera pose transformation in 3D space,this thesis further proposed a low-drift camera pose estimation solution based on a non-orthogonal coordinate system in 3D space.Experimental results showed that the proposed low-drift camera pose estimation solution achieved an accuracy improvement of 7.1% and 12.5% in terms of the average and root mean square error of absolute trajectory error,respectively,compared to the original system.3.To address the problem of visual SLAM systems being easily disturbed by moving objects in dynamic scenes,this thesis adopted a semantic segmentation neural network to extract regions of actively moving objects in the scene,and then further extracted the outer contours of these regions.Based on the distance between feature points and the outer contour,this thesis divided the extracted feature points into different sets based on their motion characteristics.Then,based on the ICP algorithm,the camera pose was initialized on a priori stationary feature point set.This rough initialization pose was used together with the scene flow geometric constraint and the epipolar geometric constraint to detect the motion consistency of the priori passive motion feature point set,and finally obtained the stationary feature point set,which was used for accurate camera pose tracking.Experimental results showed that the proposed solution achieved an accuracy improvement of 84.5%-98.4% in terms of the root mean square error of absolute pose error,and24.0%-65.2% in terms of the root mean square error of relative pose error on dynamic scene test sequences of the TUM dataset compared to the original system.更多还原