基于RGB-D和单目视觉的同时定位与建图算法研究

发布时间：2018-06-06 13:14

本文选题：RGB-D建图 + RGB-D　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：随着科学技术、计算机网络和硬件的进步,同时定位与建图(Simultaneous Localization and Mapping,SLAM)无疑已经成为了移动机器人智能化领域的研究热点之一,它对机器人的自主移动来说起到了至关重要的作用。SLAM可以使用很多方法实现,总体上可以分为滤波器方法和图优化方法。对于利用图像信息作为数据来源的SLAM问题被又称为视觉SLAM(Visual SLAM,VSLAM)。在动态、复杂度高和大尺度的环境下,利用视觉信息作为唯一的外部感知来源来解决SLAM问题是目前一个活跃的研究领域。针对该问题,本文的主要研究内容如下:首先,通过比较不同视觉传感器的优缺点并参考“图优化”方式构建了基于深度相机的RGB-D建图算法,并针对传统视觉特征鲁棒性、实时性和匹配精度较差的问题,提出了一种基于ORB视觉特征的RGB-D建图算法。然后将不同视觉特征(ORB、SIFT、SURF、FAST、GridFAST等)应用到RGB-D建图算法中并比较了它们对整个建图算法实时性、精度以及重定位能力的影响。实验证明,ORB特征在鲁棒性、实时性和匹配精度方面的性能都远优于其他视觉特征,基于ORB视觉特征的RGB-D建图算法在实时性、建图准确性和重定位能力方面效果更好。其次,针对传统关键帧选择算法单一、整个SLAM过程中关键帧数量激增的问题,提出了一种改进的关键帧选择算法,并以此为基础构建了基于ORB特征的RGB-D SLAM算法。改进的关键帧选择算法不仅整合了帧间相对运动距离、特征点跟踪以及最小视觉变化来选择关键帧,同时检测冗余关键帧并将其删除。通过在RGB-D数据集上的实验表明,改进的关键帧选择算法能够更精准、及时地选择关键帧,并在减少RGB-D SLAM算法中冗余关键帧的同时提高RGB-D SLAM算法的实时性和建图、定位精度。最后,针对RGB-D相机使用灵活性较低、特征点法鲁棒性较差的问题,利用直接法实现单目相机下的同时定位与建图。该算法使用一般的单目相机为传感器获取环境信息,克服了深度相机只能在室内环境下使用的局限性,也能在室外环境下使用。相较于特征点法,直接法直接对图像的像素灰度进行操作,能够更加充分地使用环境中丰富的信息,在显著特征不明显时也能很好地估计深度。实验结果表明,在特征不明显的情况下直接法仍适用。同时,采用直接法的MonoSLAM算法,在室内外环境下都能够快速精准的定位与建图。
[Abstract]:With the progress of science and technology, computer network and hardware, simultaneous location and mapping has undoubtedly become one of the hot research topics in the field of intelligent mobile robot. It plays an important role in autonomous movement of robot. SLAM can be realized by many methods, which can be divided into filter method and graph optimization method. For the SLAM problem which uses image information as the data source, it is also called visual SLAM(Visual slam / VSLAMN. In dynamic, high complexity and large scale environments, the use of visual information as the sole source of external perception to solve the SLAM problem is an active research field. The main research contents of this paper are as follows: firstly, by comparing the advantages and disadvantages of different vision sensors and referring to the "graph optimization" method, we construct a RGB-D mapping algorithm based on depth camera, and aim at the robustness of traditional visual features. In this paper, a RGB-D mapping algorithm based on ORB visual features is proposed for the problems of poor real time performance and poor matching accuracy. Then, different visual features are applied to the RGB-D algorithm and their effects on the real-time, accuracy and relocation ability of the whole algorithm are compared. Experiments show that Orb features are far superior to other visual features in robustness, real-time performance and matching accuracy. The RGB-D mapping algorithm based on ORB visual features is more effective in real-time, mapping accuracy and repositioning ability. Secondly, aiming at the problem that the traditional key frame selection algorithm is single and the number of key frames increases rapidly in the whole SLAM process, an improved key-frame selection algorithm is proposed, and a RGB-D SLAM algorithm based on ORB features is constructed. The improved key-frame selection algorithm not only integrates the relative motion distance between frames, feature point tracking and minimum visual changes to select key frames, but also detects redundant key frames and removes them. The experiments on RGB-D dataset show that the improved key-frame selection algorithm can select key-frame more accurately and timely, and improve the real-time performance and mapping accuracy of RGB-D SLAM algorithm while reducing redundant key-frame in RGB-D SLAM algorithm. Finally, aiming at the problem that the RGB-D camera is less flexible in use and the robustness of the feature point method is poor, the direct method is used to realize simultaneous location and map building under the single camera. The algorithm uses the general monocular camera to obtain the environmental information for the sensor, which overcomes the limitation that the depth camera can only be used in the indoor environment, and can also be used in the outdoor environment. Compared with the feature point method, the direct method can directly operate the pixel grayscale of the image, which can make full use of the rich information in the environment, and can estimate the depth well when the salient features are not obvious. The experimental results show that the direct method is still applicable when the characteristics are not obvious. At the same time, the MonoSLAM algorithm of direct method can locate and build the map quickly and accurately in indoor and outdoor environment.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】